6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciObjArray.hpp"
27 #include "ci/ciUtilities.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subtypenode.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "utilities/bitMap.inline.hpp"
50 #include "utilities/growableArray.hpp"
51 #include "utilities/powerOfTwo.hpp"
52
53 //----------------------------GraphKit-----------------------------------------
54 // Main utility constructor.
55 GraphKit::GraphKit(JVMState* jvms)
56 : Phase(Phase::Parser),
57 _env(C->env()),
58 _gvn(*C->initial_gvn()),
59 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
60 {
61 _exceptions = jvms->map()->next_exception();
62 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
63 set_jvms(jvms);
64 }
65
66 // Private constructor for parser.
67 GraphKit::GraphKit()
68 : Phase(Phase::Parser),
69 _env(C->env()),
70 _gvn(*C->initial_gvn()),
71 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
72 {
73 _exceptions = nullptr;
74 set_map(nullptr);
75 DEBUG_ONLY(_sp = -99);
76 DEBUG_ONLY(set_bci(-99));
77 }
78
79
80
81 //---------------------------clean_stack---------------------------------------
82 // Clear away rubbish from the stack area of the JVM state.
83 // This destroys any arguments that may be waiting on the stack.
328 }
329 static inline void add_one_req(Node* dstphi, Node* src) {
330 assert(is_hidden_merge(dstphi), "must be a special merge node");
331 assert(!is_hidden_merge(src), "must not be a special merge node");
332 dstphi->add_req(src);
333 }
334
335 //-----------------------combine_exception_states------------------------------
336 // This helper function combines exception states by building phis on a
337 // specially marked state-merging region. These regions and phis are
338 // untransformed, and can build up gradually. The region is marked by
339 // having a control input of its exception map, rather than null. Such
340 // regions do not appear except in this function, and in use_exception_state.
341 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
342 if (failing_internal()) {
343 return; // dying anyway...
344 }
345 JVMState* ex_jvms = ex_map->_jvms;
346 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
347 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
348 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
349 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
350 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
351 assert(ex_map->req() == phi_map->req(), "matching maps");
352 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
353 Node* hidden_merge_mark = root();
354 Node* region = phi_map->control();
355 MergeMemNode* phi_mem = phi_map->merged_memory();
356 MergeMemNode* ex_mem = ex_map->merged_memory();
357 if (region->in(0) != hidden_merge_mark) {
358 // The control input is not (yet) a specially-marked region in phi_map.
359 // Make it so, and build some phis.
360 region = new RegionNode(2);
361 _gvn.set_type(region, Type::CONTROL);
362 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
363 region->init_req(1, phi_map->control());
364 phi_map->set_control(region);
365 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
366 record_for_igvn(io_phi);
367 _gvn.set_type(io_phi, Type::ABIO);
368 phi_map->set_i_o(io_phi);
856 if (PrintMiscellaneous && (Verbose || WizardMode)) {
857 tty->print_cr("Zombie local %d: ", local);
858 jvms->dump();
859 }
860 return false;
861 }
862 }
863 }
864 return true;
865 }
866
867 #endif //ASSERT
868
869 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
870 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
871 ciMethod* cur_method = jvms->method();
872 int cur_bci = jvms->bci();
873 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
874 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
875 return Interpreter::bytecode_should_reexecute(code) ||
876 (is_anewarray && code == Bytecodes::_multianewarray);
877 // Reexecute _multianewarray bytecode which was replaced with
878 // sequence of [a]newarray. See Parse::do_multianewarray().
879 //
880 // Note: interpreter should not have it set since this optimization
881 // is limited by dimensions and guarded by flag so in some cases
882 // multianewarray() runtime calls will be generated and
883 // the bytecode should not be reexecutes (stack will not be reset).
884 } else {
885 return false;
886 }
887 }
888
889 // Helper function for adding JVMState and debug information to node
890 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
891 // Add the safepoint edges to the call (or other safepoint).
892
893 // Make sure dead locals are set to top. This
894 // should help register allocation time and cut down on the size
895 // of the deoptimization information.
896 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
924
925 if (env()->should_retain_local_variables()) {
926 // At any safepoint, this method can get breakpointed, which would
927 // then require an immediate deoptimization.
928 can_prune_locals = false; // do not prune locals
929 stack_slots_not_pruned = 0;
930 }
931
932 // do not scribble on the input jvms
933 JVMState* out_jvms = youngest_jvms->clone_deep(C);
934 call->set_jvms(out_jvms); // Start jvms list for call node
935
936 // For a known set of bytecodes, the interpreter should reexecute them if
937 // deoptimization happens. We set the reexecute state for them here
938 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
939 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
940 #ifdef ASSERT
941 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
942 assert(method() == youngest_jvms->method(), "sanity");
943 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
944 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
945 #endif // ASSERT
946 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
947 }
948
949 // Presize the call:
950 DEBUG_ONLY(uint non_debug_edges = call->req());
951 call->add_req_batch(top(), youngest_jvms->debug_depth());
952 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
953
954 // Set up edges so that the call looks like this:
955 // Call [state:] ctl io mem fptr retadr
956 // [parms:] parm0 ... parmN
957 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
958 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
959 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
960 // Note that caller debug info precedes callee debug info.
961
962 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
963 uint debug_ptr = call->req();
964
965 // Loop over the map input edges associated with jvms, add them
966 // to the call node, & reset all offsets to match call node array.
967 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
968 uint debug_end = debug_ptr;
969 uint debug_start = debug_ptr - in_jvms->debug_size();
970 debug_ptr = debug_start; // back up the ptr
971
972 uint p = debug_start; // walks forward in [debug_start, debug_end)
973 uint j, k, l;
974 SafePointNode* in_map = in_jvms->map();
975 out_jvms->set_map(call);
976
977 if (can_prune_locals) {
978 assert(in_jvms->method() == out_jvms->method(), "sanity");
979 // If the current throw can reach an exception handler in this JVMS,
980 // then we must keep everything live that can reach that handler.
981 // As a quick and dirty approximation, we look for any handlers at all.
982 if (in_jvms->method()->has_exception_handlers()) {
983 can_prune_locals = false;
984 }
985 }
986
987 // Add the Locals
988 k = in_jvms->locoff();
989 l = in_jvms->loc_size();
990 out_jvms->set_locoff(p);
991 if (!can_prune_locals) {
992 for (j = 0; j < l; j++)
993 call->set_req(p++, in_map->in(k+j));
994 } else {
995 p += l; // already set to top above by add_req_batch
996 }
997
998 // Add the Expression Stack
999 k = in_jvms->stkoff();
1000 l = in_jvms->sp();
1001 out_jvms->set_stkoff(p);
1002 if (!can_prune_locals) {
1003 for (j = 0; j < l; j++)
1004 call->set_req(p++, in_map->in(k+j));
1005 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1006 // Divide stack into {S0,...,S1}, where S0 is set to top.
1007 uint s1 = stack_slots_not_pruned;
1008 stack_slots_not_pruned = 0; // for next iteration
1009 if (s1 > l) s1 = l;
1010 uint s0 = l - s1;
1011 p += s0; // skip the tops preinstalled by add_req_batch
1012 for (j = s0; j < l; j++)
1013 call->set_req(p++, in_map->in(k+j));
1014 } else {
1015 p += l; // already set to top above by add_req_batch
1016 }
1017
1018 // Add the Monitors
1019 k = in_jvms->monoff();
1020 l = in_jvms->mon_size();
1021 out_jvms->set_monoff(p);
1022 for (j = 0; j < l; j++)
1023 call->set_req(p++, in_map->in(k+j));
1024
1025 // Copy any scalar object fields.
1026 k = in_jvms->scloff();
1027 l = in_jvms->scl_size();
1028 out_jvms->set_scloff(p);
1029 for (j = 0; j < l; j++)
1030 call->set_req(p++, in_map->in(k+j));
1031
1032 // Finish the new jvms.
1033 out_jvms->set_endoff(p);
1034
1035 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1036 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1037 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1038 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1039 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1040 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1041
1042 // Update the two tail pointers in parallel.
1043 out_jvms = out_jvms->caller();
1044 in_jvms = in_jvms->caller();
1045 }
1046
1047 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1048
1049 // Test the correctness of JVMState::debug_xxx accessors:
1050 assert(call->jvms()->debug_start() == non_debug_edges, "");
1051 assert(call->jvms()->debug_end() == call->req(), "");
1052 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1053 }
1054
1055 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1056 Bytecodes::Code code = java_bc();
1057 if (code == Bytecodes::_wide) {
1058 code = method()->java_code_at_bci(bci() + 1);
1059 }
1060
1061 if (code != Bytecodes::_illegal) {
1062 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1198 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1199 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1200 return _gvn.transform( new AndLNode(conv, mask) );
1201 }
1202
1203 Node* GraphKit::ConvL2I(Node* offset) {
1204 // short-circuit a common case
1205 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1206 if (offset_con != (jlong)Type::OffsetBot) {
1207 return intcon((int) offset_con);
1208 }
1209 return _gvn.transform( new ConvL2INode(offset));
1210 }
1211
1212 //-------------------------load_object_klass-----------------------------------
1213 Node* GraphKit::load_object_klass(Node* obj) {
1214 // Special-case a fresh allocation to avoid building nodes:
1215 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1216 if (akls != nullptr) return akls;
1217 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1218 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1219 }
1220
1221 //-------------------------load_array_length-----------------------------------
1222 Node* GraphKit::load_array_length(Node* array) {
1223 // Special-case a fresh allocation to avoid building nodes:
1224 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1225 Node *alen;
1226 if (alloc == nullptr) {
1227 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1228 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1229 } else {
1230 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1231 }
1232 return alen;
1233 }
1234
1235 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1236 const TypeOopPtr* oop_type,
1237 bool replace_length_in_map) {
1238 Node* length = alloc->Ideal_length();
1247 replace_in_map(length, ccast);
1248 }
1249 return ccast;
1250 }
1251 }
1252 return length;
1253 }
1254
1255 //------------------------------do_null_check----------------------------------
1256 // Helper function to do a null pointer check. Returned value is
1257 // the incoming address with null casted away. You are allowed to use the
1258 // not-null value only if you are control dependent on the test.
1259 #ifndef PRODUCT
1260 extern uint explicit_null_checks_inserted,
1261 explicit_null_checks_elided;
1262 #endif
1263 Node* GraphKit::null_check_common(Node* value, BasicType type,
1264 // optional arguments for variations:
1265 bool assert_null,
1266 Node* *null_control,
1267 bool speculative) {
1268 assert(!assert_null || null_control == nullptr, "not both at once");
1269 if (stopped()) return top();
1270 NOT_PRODUCT(explicit_null_checks_inserted++);
1271
1272 // Construct null check
1273 Node *chk = nullptr;
1274 switch(type) {
1275 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1276 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1277 case T_ARRAY : // fall through
1278 type = T_OBJECT; // simplify further tests
1279 case T_OBJECT : {
1280 const Type *t = _gvn.type( value );
1281
1282 const TypeOopPtr* tp = t->isa_oopptr();
1283 if (tp != nullptr && !tp->is_loaded()
1284 // Only for do_null_check, not any of its siblings:
1285 && !assert_null && null_control == nullptr) {
1286 // Usually, any field access or invocation on an unloaded oop type
1287 // will simply fail to link, since the statically linked class is
1288 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1289 // the static class is loaded but the sharper oop type is not.
1290 // Rather than checking for this obscure case in lots of places,
1291 // we simply observe that a null check on an unloaded class
1355 }
1356 Node *oldcontrol = control();
1357 set_control(cfg);
1358 Node *res = cast_not_null(value);
1359 set_control(oldcontrol);
1360 NOT_PRODUCT(explicit_null_checks_elided++);
1361 return res;
1362 }
1363 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1364 if (cfg == nullptr) break; // Quit at region nodes
1365 depth++;
1366 }
1367 }
1368
1369 //-----------
1370 // Branch to failure if null
1371 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1372 Deoptimization::DeoptReason reason;
1373 if (assert_null) {
1374 reason = Deoptimization::reason_null_assert(speculative);
1375 } else if (type == T_OBJECT) {
1376 reason = Deoptimization::reason_null_check(speculative);
1377 } else {
1378 reason = Deoptimization::Reason_div0_check;
1379 }
1380 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1381 // ciMethodData::has_trap_at will return a conservative -1 if any
1382 // must-be-null assertion has failed. This could cause performance
1383 // problems for a method after its first do_null_assert failure.
1384 // Consider using 'Reason_class_check' instead?
1385
1386 // To cause an implicit null check, we set the not-null probability
1387 // to the maximum (PROB_MAX). For an explicit check the probability
1388 // is set to a smaller value.
1389 if (null_control != nullptr || too_many_traps(reason)) {
1390 // probability is less likely
1391 ok_prob = PROB_LIKELY_MAG(3);
1392 } else if (!assert_null &&
1393 (ImplicitNullCheckThreshold > 0) &&
1394 method() != nullptr &&
1395 (method()->method_data()->trap_count(reason)
1429 }
1430
1431 if (assert_null) {
1432 // Cast obj to null on this path.
1433 replace_in_map(value, zerocon(type));
1434 return zerocon(type);
1435 }
1436
1437 // Cast obj to not-null on this path, if there is no null_control.
1438 // (If there is a null_control, a non-null value may come back to haunt us.)
1439 if (type == T_OBJECT) {
1440 Node* cast = cast_not_null(value, false);
1441 if (null_control == nullptr || (*null_control) == top())
1442 replace_in_map(value, cast);
1443 value = cast;
1444 }
1445
1446 return value;
1447 }
1448
1449
1450 //------------------------------cast_not_null----------------------------------
1451 // Cast obj to not-null on this path
1452 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1453 const Type *t = _gvn.type(obj);
1454 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1455 // Object is already not-null?
1456 if( t == t_not_null ) return obj;
1457
1458 Node* cast = new CastPPNode(control(), obj,t_not_null);
1459 cast = _gvn.transform( cast );
1460
1461 // Scan for instances of 'obj' in the current JVM mapping.
1462 // These instances are known to be not-null after the test.
1463 if (do_replace_in_map)
1464 replace_in_map(obj, cast);
1465
1466 return cast; // Return casted value
1467 }
1468
1469 // Sometimes in intrinsics, we implicitly know an object is not null
1470 // (there's no actual null check) so we can cast it to not null. In
1471 // the course of optimizations, the input to the cast can become null.
1472 // In that case that data path will die and we need the control path
1473 // to become dead as well to keep the graph consistent. So we have to
1474 // add a check for null for which one branch can't be taken. It uses
1475 // an OpaqueNotNull node that will cause the check to be removed after loop
1476 // opts so the test goes away and the compiled code doesn't execute a
1477 // useless check.
1478 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1479 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1480 return value;
1481 }
1482 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1483 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1484 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1485 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1486 _gvn.set_type(iff, iff->Value(&_gvn));
1487 if (!tst->is_Con()) {
1488 record_for_igvn(iff);
1560 // These are layered on top of the factory methods in LoadNode and StoreNode,
1561 // and integrate with the parser's memory state and _gvn engine.
1562 //
1563
1564 // factory methods in "int adr_idx"
1565 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1566 MemNode::MemOrd mo,
1567 LoadNode::ControlDependency control_dependency,
1568 bool require_atomic_access,
1569 bool unaligned,
1570 bool mismatched,
1571 bool unsafe,
1572 uint8_t barrier_data) {
1573 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1574 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1575 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1576 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1577 Node* mem = memory(adr_idx);
1578 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1579 ld = _gvn.transform(ld);
1580 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1581 // Improve graph before escape analysis and boxing elimination.
1582 record_for_igvn(ld);
1583 if (ld->is_DecodeN()) {
1584 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1585 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1586 // a Phi). Recording such cases is still perfectly sound, but may be
1587 // unnecessary and result in some minor IGVN overhead.
1588 record_for_igvn(ld->in(1));
1589 }
1590 }
1591 return ld;
1592 }
1593
1594 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1595 MemNode::MemOrd mo,
1596 bool require_atomic_access,
1597 bool unaligned,
1598 bool mismatched,
1599 bool unsafe,
1613 if (unsafe) {
1614 st->as_Store()->set_unsafe_access();
1615 }
1616 st->as_Store()->set_barrier_data(barrier_data);
1617 st = _gvn.transform(st);
1618 set_memory(st, adr_idx);
1619 // Back-to-back stores can only remove intermediate store with DU info
1620 // so push on worklist for optimizer.
1621 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1622 record_for_igvn(st);
1623
1624 return st;
1625 }
1626
1627 Node* GraphKit::access_store_at(Node* obj,
1628 Node* adr,
1629 const TypePtr* adr_type,
1630 Node* val,
1631 const Type* val_type,
1632 BasicType bt,
1633 DecoratorSet decorators) {
1634 // Transformation of a value which could be null pointer (CastPP #null)
1635 // could be delayed during Parse (for example, in adjust_map_after_if()).
1636 // Execute transformation here to avoid barrier generation in such case.
1637 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1638 val = _gvn.makecon(TypePtr::NULL_PTR);
1639 }
1640
1641 if (stopped()) {
1642 return top(); // Dead path ?
1643 }
1644
1645 assert(val != nullptr, "not dead path");
1646
1647 C2AccessValuePtr addr(adr, adr_type);
1648 C2AccessValue value(val, val_type);
1649 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1650 if (access.is_raw()) {
1651 return _barrier_set->BarrierSetC2::store_at(access, value);
1652 } else {
1653 return _barrier_set->store_at(access, value);
1654 }
1655 }
1656
1657 Node* GraphKit::access_load_at(Node* obj, // containing obj
1658 Node* adr, // actual address to store val at
1659 const TypePtr* adr_type,
1660 const Type* val_type,
1661 BasicType bt,
1662 DecoratorSet decorators) {
1663 if (stopped()) {
1664 return top(); // Dead path ?
1665 }
1666
1667 C2AccessValuePtr addr(adr, adr_type);
1668 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1669 if (access.is_raw()) {
1670 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1671 } else {
1672 return _barrier_set->load_at(access, val_type);
1673 }
1674 }
1675
1676 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1677 const Type* val_type,
1678 BasicType bt,
1679 DecoratorSet decorators) {
1680 if (stopped()) {
1681 return top(); // Dead path ?
1682 }
1683
1684 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1685 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1686 if (access.is_raw()) {
1687 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1688 } else {
1753 Node* new_val,
1754 const Type* value_type,
1755 BasicType bt,
1756 DecoratorSet decorators) {
1757 C2AccessValuePtr addr(adr, adr_type);
1758 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1759 if (access.is_raw()) {
1760 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1761 } else {
1762 return _barrier_set->atomic_add_at(access, new_val, value_type);
1763 }
1764 }
1765
1766 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1767 return _barrier_set->clone(this, src, dst, size, is_array);
1768 }
1769
1770 //-------------------------array_element_address-------------------------
1771 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1772 const TypeInt* sizetype, Node* ctrl) {
1773 uint shift = exact_log2(type2aelembytes(elembt));
1774 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1775
1776 // short-circuit a common case (saves lots of confusing waste motion)
1777 jint idx_con = find_int_con(idx, -1);
1778 if (idx_con >= 0) {
1779 intptr_t offset = header + ((intptr_t)idx_con << shift);
1780 return basic_plus_adr(ary, offset);
1781 }
1782
1783 // must be correct type for alignment purposes
1784 Node* base = basic_plus_adr(ary, header);
1785 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1786 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1787 return basic_plus_adr(ary, base, scale);
1788 }
1789
1790 //-------------------------load_array_element-------------------------
1791 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1792 const Type* elemtype = arytype->elem();
1793 BasicType elembt = elemtype->array_element_basic_type();
1794 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1795 if (elembt == T_NARROWOOP) {
1796 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1797 }
1798 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1799 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1800 return ld;
1801 }
1802
1803 //-------------------------set_arguments_for_java_call-------------------------
1804 // Arguments (pre-popped from the stack) are taken from the JVMS.
1805 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1806 // Add the call arguments:
1807 uint nargs = call->method()->arg_size();
1808 for (uint i = 0; i < nargs; i++) {
1809 Node* arg = argument(i);
1810 call->init_req(i + TypeFunc::Parms, arg);
1811 }
1812 }
1813
1814 //---------------------------set_edges_for_java_call---------------------------
1815 // Connect a newly created call into the current JVMS.
1816 // A return value node (if any) is returned from set_edges_for_java_call.
1817 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1818
1819 // Add the predefined inputs:
1820 call->init_req( TypeFunc::Control, control() );
1821 call->init_req( TypeFunc::I_O , i_o() );
1822 call->init_req( TypeFunc::Memory , reset_memory() );
1823 call->init_req( TypeFunc::FramePtr, frameptr() );
1824 call->init_req( TypeFunc::ReturnAdr, top() );
1825
1826 add_safepoint_edges(call, must_throw);
1827
1828 Node* xcall = _gvn.transform(call);
1829
1830 if (xcall == top()) {
1831 set_control(top());
1832 return;
1833 }
1834 assert(xcall == call, "call identity is stable");
1835
1836 // Re-use the current map to produce the result.
1837
1838 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1839 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1840 set_all_memory_call(xcall, separate_io_proj);
1841
1842 //return xcall; // no need, caller already has it
1843 }
1844
1845 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1846 if (stopped()) return top(); // maybe the call folded up?
1847
1848 // Capture the return value, if any.
1849 Node* ret;
1850 if (call->method() == nullptr ||
1851 call->method()->return_type()->basic_type() == T_VOID)
1852 ret = top();
1853 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1854
1855 // Note: Since any out-of-line call can produce an exception,
1856 // we always insert an I_O projection from the call into the result.
1857
1858 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1859
1860 if (separate_io_proj) {
1861 // The caller requested separate projections be used by the fall
1862 // through and exceptional paths, so replace the projections for
1863 // the fall through path.
1864 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1865 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1866 }
1867 return ret;
1868 }
1869
1870 //--------------------set_predefined_input_for_runtime_call--------------------
1871 // Reading and setting the memory state is way conservative here.
1872 // The real problem is that I am not doing real Type analysis on memory,
1873 // so I cannot distinguish card mark stores from other stores. Across a GC
1874 // point the Store Barrier and the card mark memory has to agree. I cannot
1875 // have a card mark store and its barrier split across the GC point from
1876 // either above or below. Here I get that to happen by reading ALL of memory.
1877 // A better answer would be to separate out card marks from other memory.
1878 // For now, return the input memory state, so that it can be reused
1879 // after the call, if this call has restricted memory effects.
1880 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1881 // Set fixed predefined input arguments
1882 call->init_req(TypeFunc::Control, control());
1883 call->init_req(TypeFunc::I_O, top()); // does no i/o
1884 call->init_req(TypeFunc::ReturnAdr, top());
1885 if (call->is_CallLeafPure()) {
1886 call->init_req(TypeFunc::Memory, top());
1948 if (use->is_MergeMem()) {
1949 wl.push(use);
1950 }
1951 }
1952 }
1953
1954 // Replace the call with the current state of the kit.
1955 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
1956 JVMState* ejvms = nullptr;
1957 if (has_exceptions()) {
1958 ejvms = transfer_exceptions_into_jvms();
1959 }
1960
1961 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1962 ReplacedNodes replaced_nodes_exception;
1963 Node* ex_ctl = top();
1964
1965 SafePointNode* final_state = stop();
1966
1967 // Find all the needed outputs of this call
1968 CallProjections callprojs;
1969 call->extract_projections(&callprojs, true, do_asserts);
1970
1971 Unique_Node_List wl;
1972 Node* init_mem = call->in(TypeFunc::Memory);
1973 Node* final_mem = final_state->in(TypeFunc::Memory);
1974 Node* final_ctl = final_state->in(TypeFunc::Control);
1975 Node* final_io = final_state->in(TypeFunc::I_O);
1976
1977 // Replace all the old call edges with the edges from the inlining result
1978 if (callprojs.fallthrough_catchproj != nullptr) {
1979 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1980 }
1981 if (callprojs.fallthrough_memproj != nullptr) {
1982 if (final_mem->is_MergeMem()) {
1983 // Parser's exits MergeMem was not transformed but may be optimized
1984 final_mem = _gvn.transform(final_mem);
1985 }
1986 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1987 add_mergemem_users_to_worklist(wl, final_mem);
1988 }
1989 if (callprojs.fallthrough_ioproj != nullptr) {
1990 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1991 }
1992
1993 // Replace the result with the new result if it exists and is used
1994 if (callprojs.resproj != nullptr && result != nullptr) {
1995 C->gvn_replace_by(callprojs.resproj, result);
1996 }
1997
1998 if (ejvms == nullptr) {
1999 // No exception edges to simply kill off those paths
2000 if (callprojs.catchall_catchproj != nullptr) {
2001 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
2002 }
2003 if (callprojs.catchall_memproj != nullptr) {
2004 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
2005 }
2006 if (callprojs.catchall_ioproj != nullptr) {
2007 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
2008 }
2009 // Replace the old exception object with top
2010 if (callprojs.exobj != nullptr) {
2011 C->gvn_replace_by(callprojs.exobj, C->top());
2012 }
2013 } else {
2014 GraphKit ekit(ejvms);
2015
2016 // Load my combined exception state into the kit, with all phis transformed:
2017 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2018 replaced_nodes_exception = ex_map->replaced_nodes();
2019
2020 Node* ex_oop = ekit.use_exception_state(ex_map);
2021
2022 if (callprojs.catchall_catchproj != nullptr) {
2023 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2024 ex_ctl = ekit.control();
2025 }
2026 if (callprojs.catchall_memproj != nullptr) {
2027 Node* ex_mem = ekit.reset_memory();
2028 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2029 add_mergemem_users_to_worklist(wl, ex_mem);
2030 }
2031 if (callprojs.catchall_ioproj != nullptr) {
2032 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2033 }
2034
2035 // Replace the old exception object with the newly created one
2036 if (callprojs.exobj != nullptr) {
2037 C->gvn_replace_by(callprojs.exobj, ex_oop);
2038 }
2039 }
2040
2041 // Disconnect the call from the graph
2042 call->disconnect_inputs(C);
2043 C->gvn_replace_by(call, C->top());
2044
2045 // Clean up any MergeMems that feed other MergeMems since the
2046 // optimizer doesn't like that.
2047 while (wl.size() > 0) {
2048 _gvn.transform(wl.pop());
2049 }
2050
2051 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2052 replaced_nodes.apply(C, final_ctl);
2053 }
2054 if (!ex_ctl->is_top() && do_replaced_nodes) {
2055 replaced_nodes_exception.apply(C, ex_ctl);
2056 }
2057 }
2058
2059
2060 //------------------------------increment_counter------------------------------
2061 // for statistics: increment a VM counter by 1
2062
2063 void GraphKit::increment_counter(address counter_addr) {
2064 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2065 increment_counter(adr1);
2066 }
2067
2068 void GraphKit::increment_counter(Node* counter_addr) {
2069 Node* ctrl = control();
2070 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2071 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2241 *
2242 * @param n node that the type applies to
2243 * @param exact_kls type from profiling
2244 * @param maybe_null did profiling see null?
2245 *
2246 * @return node with improved type
2247 */
2248 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2249 const Type* current_type = _gvn.type(n);
2250 assert(UseTypeSpeculation, "type speculation must be on");
2251
2252 const TypePtr* speculative = current_type->speculative();
2253
2254 // Should the klass from the profile be recorded in the speculative type?
2255 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2256 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2257 const TypeOopPtr* xtype = tklass->as_instance_type();
2258 assert(xtype->klass_is_exact(), "Should be exact");
2259 // Any reason to believe n is not null (from this profiling or a previous one)?
2260 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2261 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2262 // record the new speculative type's depth
2263 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2264 speculative = speculative->with_inline_depth(jvms()->depth());
2265 } else if (current_type->would_improve_ptr(ptr_kind)) {
2266 // Profiling report that null was never seen so we can change the
2267 // speculative type to non null ptr.
2268 if (ptr_kind == ProfileAlwaysNull) {
2269 speculative = TypePtr::NULL_PTR;
2270 } else {
2271 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2272 const TypePtr* ptr = TypePtr::NOTNULL;
2273 if (speculative != nullptr) {
2274 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2275 } else {
2276 speculative = ptr;
2277 }
2278 }
2279 }
2280
2281 if (speculative != current_type->speculative()) {
2282 // Build a type with a speculative type (what we think we know
2283 // about the type but will need a guard when we use it)
2284 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2285 // We're changing the type, we need a new CheckCast node to carry
2286 // the new type. The new type depends on the control: what
2287 // profiling tells us is only valid from here as far as we can
2288 // tell.
2289 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2290 cast = _gvn.transform(cast);
2291 replace_in_map(n, cast);
2292 n = cast;
2293 }
2294
2295 return n;
2296 }
2297
2298 /**
2299 * Record profiling data from receiver profiling at an invoke with the
2300 * type system so that it can propagate it (speculation)
2301 *
2302 * @param n receiver node
2303 *
2304 * @return node with improved type
2305 */
2306 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2307 if (!UseTypeSpeculation) {
2308 return n;
2309 }
2310 ciKlass* exact_kls = profile_has_unique_klass();
2311 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2312 if ((java_bc() == Bytecodes::_checkcast ||
2313 java_bc() == Bytecodes::_instanceof ||
2314 java_bc() == Bytecodes::_aastore) &&
2315 method()->method_data()->is_mature()) {
2316 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2317 if (data != nullptr) {
2318 if (!data->as_BitData()->null_seen()) {
2319 ptr_kind = ProfileNeverNull;
2320 } else {
2321 if (TypeProfileCasts) {
2322 assert(data->is_ReceiverTypeData(), "bad profile data type");
2323 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2324 uint i = 0;
2325 for (; i < call->row_limit(); i++) {
2326 ciKlass* receiver = call->receiver(i);
2327 if (receiver != nullptr) {
2328 break;
2329 }
2330 }
2331 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2332 }
2333 }
2334 }
2335 }
2336 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2337 }
2338
2339 /**
2340 * Record profiling data from argument profiling at an invoke with the
2341 * type system so that it can propagate it (speculation)
2342 *
2343 * @param dest_method target method for the call
2344 * @param bc what invoke bytecode is this?
2345 */
2346 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2347 if (!UseTypeSpeculation) {
2348 return;
2349 }
2350 const TypeFunc* tf = TypeFunc::make(dest_method);
2351 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2352 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2353 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2354 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2355 if (is_reference_type(targ->basic_type())) {
2356 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2357 ciKlass* better_type = nullptr;
2358 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2359 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2360 }
2361 i++;
2362 }
2363 }
2364 }
2365
2366 /**
2367 * Record profiling data from parameter profiling at an invoke with
2368 * the type system so that it can propagate it (speculation)
2369 */
2370 void GraphKit::record_profiled_parameters_for_speculation() {
2371 if (!UseTypeSpeculation) {
2372 return;
2373 }
2374 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2494 // The first null ends the list.
2495 Node* parm0, Node* parm1,
2496 Node* parm2, Node* parm3,
2497 Node* parm4, Node* parm5,
2498 Node* parm6, Node* parm7) {
2499 assert(call_addr != nullptr, "must not call null targets");
2500
2501 // Slow-path call
2502 bool is_leaf = !(flags & RC_NO_LEAF);
2503 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2504 if (call_name == nullptr) {
2505 assert(!is_leaf, "must supply name for leaf");
2506 call_name = OptoRuntime::stub_name(call_addr);
2507 }
2508 CallNode* call;
2509 if (!is_leaf) {
2510 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2511 } else if (flags & RC_NO_FP) {
2512 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2513 } else if (flags & RC_VECTOR){
2514 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2515 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2516 } else if (flags & RC_PURE) {
2517 assert(adr_type == nullptr, "pure call does not touch memory");
2518 call = new CallLeafPureNode(call_type, call_addr, call_name);
2519 } else {
2520 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2521 }
2522
2523 // The following is similar to set_edges_for_java_call,
2524 // except that the memory effects of the call are restricted to AliasIdxRaw.
2525
2526 // Slow path call has no side-effects, uses few values
2527 bool wide_in = !(flags & RC_NARROW_MEM);
2528 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2529
2530 Node* prev_mem = nullptr;
2531 if (wide_in) {
2532 prev_mem = set_predefined_input_for_runtime_call(call);
2533 } else {
2534 assert(!wide_out, "narrow in => narrow out");
2535 Node* narrow_mem = memory(adr_type);
2536 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2537 }
2538
2539 // Hook each parm in order. Stop looking at the first null.
2540 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2541 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2542 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2543 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2544 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2545 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2546 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2547 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2548 /* close each nested if ===> */ } } } } } } } }
2549 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2550
2551 if (!is_leaf) {
2552 // Non-leaves can block and take safepoints:
2553 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2554 }
2555 // Non-leaves can throw exceptions:
2556 if (has_io) {
2557 call->set_req(TypeFunc::I_O, i_o());
2558 }
2559
2560 if (flags & RC_UNCOMMON) {
2561 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2562 // (An "if" probability corresponds roughly to an unconditional count.
2563 // Sort of.)
2564 call->set_cnt(PROB_UNLIKELY_MAG(4));
2565 }
2566
2567 Node* c = _gvn.transform(call);
2568 assert(c == call, "cannot disappear");
2569
2577
2578 if (has_io) {
2579 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2580 }
2581 return call;
2582
2583 }
2584
2585 // i2b
2586 Node* GraphKit::sign_extend_byte(Node* in) {
2587 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2588 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2589 }
2590
2591 // i2s
2592 Node* GraphKit::sign_extend_short(Node* in) {
2593 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2594 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2595 }
2596
2597 //------------------------------merge_memory-----------------------------------
2598 // Merge memory from one path into the current memory state.
2599 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2600 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2601 Node* old_slice = mms.force_memory();
2602 Node* new_slice = mms.memory2();
2603 if (old_slice != new_slice) {
2604 PhiNode* phi;
2605 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2606 if (mms.is_empty()) {
2607 // clone base memory Phi's inputs for this memory slice
2608 assert(old_slice == mms.base_memory(), "sanity");
2609 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2610 _gvn.set_type(phi, Type::MEMORY);
2611 for (uint i = 1; i < phi->req(); i++) {
2612 phi->init_req(i, old_slice->in(i));
2613 }
2614 } else {
2615 phi = old_slice->as_Phi(); // Phi was generated already
2616 }
2673 gvn.transform(iff);
2674 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2675 return iff;
2676 }
2677
2678 //-------------------------------gen_subtype_check-----------------------------
2679 // Generate a subtyping check. Takes as input the subtype and supertype.
2680 // Returns 2 values: sets the default control() to the true path and returns
2681 // the false path. Only reads invariant memory; sets no (visible) memory.
2682 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2683 // but that's not exposed to the optimizer. This call also doesn't take in an
2684 // Object; if you wish to check an Object you need to load the Object's class
2685 // prior to coming here.
2686 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2687 ciMethod* method, int bci) {
2688 Compile* C = gvn.C;
2689 if ((*ctrl)->is_top()) {
2690 return C->top();
2691 }
2692
2693 // Fast check for identical types, perhaps identical constants.
2694 // The types can even be identical non-constants, in cases
2695 // involving Array.newInstance, Object.clone, etc.
2696 if (subklass == superklass)
2697 return C->top(); // false path is dead; no test needed.
2698
2699 if (gvn.type(superklass)->singleton()) {
2700 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2701 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2702
2703 // In the common case of an exact superklass, try to fold up the
2704 // test before generating code. You may ask, why not just generate
2705 // the code and then let it fold up? The answer is that the generated
2706 // code will necessarily include null checks, which do not always
2707 // completely fold away. If they are also needless, then they turn
2708 // into a performance loss. Example:
2709 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2710 // Here, the type of 'fa' is often exact, so the store check
2711 // of fa[1]=x will fold up, without testing the nullness of x.
2712 //
2713 // At macro expansion, we would have already folded the SubTypeCheckNode
2714 // being expanded here because we always perform the static sub type
2715 // check in SubTypeCheckNode::sub() regardless of whether
2716 // StressReflectiveCode is set or not. We can therefore skip this
2717 // static check when StressReflectiveCode is on.
2718 switch (C->static_subtype_check(superk, subk)) {
2719 case Compile::SSC_always_false:
2720 {
2721 Node* always_fail = *ctrl;
2722 *ctrl = gvn.C->top();
2723 return always_fail;
2724 }
2725 case Compile::SSC_always_true:
2726 return C->top();
2727 case Compile::SSC_easy_test:
2728 {
2729 // Just do a direct pointer compare and be done.
2730 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2731 *ctrl = gvn.transform(new IfTrueNode(iff));
2732 return gvn.transform(new IfFalseNode(iff));
2733 }
2734 case Compile::SSC_full_test:
2735 break;
2736 default:
2737 ShouldNotReachHere();
2738 }
2739 }
2740
2741 // %%% Possible further optimization: Even if the superklass is not exact,
2742 // if the subklass is the unique subtype of the superklass, the check
2743 // will always succeed. We could leave a dependency behind to ensure this.
2744
2745 // First load the super-klass's check-offset
2746 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2747 Node* m = C->immutable_memory();
2748 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2749 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2750 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2788 gvn.record_for_igvn(r_ok_subtype);
2789
2790 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2791 // SubTypeCheck node
2792 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2793 ciCallProfile profile = method->call_profile_at_bci(bci);
2794 float total_prob = 0;
2795 for (int i = 0; profile.has_receiver(i); ++i) {
2796 float prob = profile.receiver_prob(i);
2797 total_prob += prob;
2798 }
2799 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2800 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2801 for (int i = 0; profile.has_receiver(i); ++i) {
2802 ciKlass* klass = profile.receiver(i);
2803 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2804 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2805 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2806 continue;
2807 }
2808 float prob = profile.receiver_prob(i);
2809 ConNode* klass_node = gvn.makecon(klass_t);
2810 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2811 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2812
2813 if (result == Compile::SSC_always_true) {
2814 r_ok_subtype->add_req(iftrue);
2815 } else {
2816 assert(result == Compile::SSC_always_false, "");
2817 r_not_subtype->add_req(iftrue);
2818 }
2819 *ctrl = gvn.transform(new IfFalseNode(iff));
2820 }
2821 }
2822 }
2823
2824 // See if we get an immediate positive hit. Happens roughly 83% of the
2825 // time. Test to see if the value loaded just previously from the subklass
2826 // is exactly the superklass.
2827 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2841 igvn->remove_globally_dead_node(r_not_subtype);
2842 }
2843 return not_subtype_ctrl;
2844 }
2845
2846 r_ok_subtype->init_req(1, iftrue1);
2847
2848 // Check for immediate negative hit. Happens roughly 11% of the time (which
2849 // is roughly 63% of the remaining cases). Test to see if the loaded
2850 // check-offset points into the subklass display list or the 1-element
2851 // cache. If it points to the display (and NOT the cache) and the display
2852 // missed then it's not a subtype.
2853 Node *cacheoff = gvn.intcon(cacheoff_con);
2854 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2855 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2856 *ctrl = gvn.transform(new IfFalseNode(iff2));
2857
2858 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2859 // No performance impact (too rare) but allows sharing of secondary arrays
2860 // which has some footprint reduction.
2861 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2862 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2863 *ctrl = gvn.transform(new IfFalseNode(iff3));
2864
2865 // -- Roads not taken here: --
2866 // We could also have chosen to perform the self-check at the beginning
2867 // of this code sequence, as the assembler does. This would not pay off
2868 // the same way, since the optimizer, unlike the assembler, can perform
2869 // static type analysis to fold away many successful self-checks.
2870 // Non-foldable self checks work better here in second position, because
2871 // the initial primary superclass check subsumes a self-check for most
2872 // types. An exception would be a secondary type like array-of-interface,
2873 // which does not appear in its own primary supertype display.
2874 // Finally, we could have chosen to move the self-check into the
2875 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2876 // dependent manner. But it is worthwhile to have the check here,
2877 // where it can be perhaps be optimized. The cost in code space is
2878 // small (register compare, branch).
2879
2880 // Now do a linear scan of the secondary super-klass array. Again, no real
2881 // performance impact (too rare) but it's gotta be done.
2882 // Since the code is rarely used, there is no penalty for moving it
2883 // out of line, and it can only improve I-cache density.
2884 // The decision to inline or out-of-line this final check is platform
2885 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2886 Node* psc = gvn.transform(
2887 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2888
2889 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2890 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2891 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2892
2893 // Return false path; set default control to true path.
2894 *ctrl = gvn.transform(r_ok_subtype);
2895 return gvn.transform(r_not_subtype);
2896 }
2897
2898 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2899 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2900 if (expand_subtype_check) {
2901 MergeMemNode* mem = merged_memory();
2902 Node* ctrl = control();
2903 Node* subklass = obj_or_subklass;
2904 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2905 subklass = load_object_klass(obj_or_subklass);
2906 }
2907
2908 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2909 set_control(ctrl);
2910 return n;
2911 }
2912
2913 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2914 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2915 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2916 set_control(_gvn.transform(new IfTrueNode(iff)));
2917 return _gvn.transform(new IfFalseNode(iff));
2918 }
2919
2920 // Profile-driven exact type check:
2921 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2922 float prob,
2923 Node* *casted_receiver) {
2924 assert(!klass->is_interface(), "no exact type check on interfaces");
2925
2926 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
2927 Node* recv_klass = load_object_klass(receiver);
2928 Node* want_klass = makecon(tklass);
2929 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
2930 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
2931 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2932 set_control( _gvn.transform(new IfTrueNode (iff)));
2933 Node* fail = _gvn.transform(new IfFalseNode(iff));
2934
2935 if (!stopped()) {
2936 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2937 const TypeOopPtr* recvx_type = tklass->as_instance_type();
2938 assert(recvx_type->klass_is_exact(), "");
2939
2940 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
2941 // Subsume downstream occurrences of receiver with a cast to
2942 // recv_xtype, since now we know what the type will be.
2943 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
2944 (*casted_receiver) = _gvn.transform(cast);
2945 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
2946 // (User must make the replace_in_map call.)
2947 }
2948 }
2949
2950 return fail;
2951 }
2952
2953 //------------------------------subtype_check_receiver-------------------------
2954 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2955 Node** casted_receiver) {
2956 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
2957 Node* want_klass = makecon(tklass);
2958
2959 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2960
2961 // Ignore interface type information until interface types are properly tracked.
2962 if (!stopped() && !klass->is_interface()) {
2963 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2964 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2965 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
2966 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2967 (*casted_receiver) = _gvn.transform(cast);
2968 }
2969 }
2970
2971 return slow_ctl;
2972 }
2973
2974 //------------------------------seems_never_null-------------------------------
2975 // Use null_seen information if it is available from the profile.
2976 // If we see an unexpected null at a type check we record it and force a
2977 // recompile; the offending check will be recompiled to handle nulls.
2978 // If we see several offending BCIs, then all checks in the
2979 // method will be recompiled.
2980 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2981 speculating = !_gvn.type(obj)->speculative_maybe_null();
2982 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2983 if (UncommonNullCast // Cutout for this technique
2984 && obj != null() // And not the -Xcomp stupid case?
2985 && !too_many_traps(reason)
2986 ) {
2987 if (speculating) {
3056
3057 //------------------------maybe_cast_profiled_receiver-------------------------
3058 // If the profile has seen exactly one type, narrow to exactly that type.
3059 // Subsequent type checks will always fold up.
3060 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3061 const TypeKlassPtr* require_klass,
3062 ciKlass* spec_klass,
3063 bool safe_for_replace) {
3064 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3065
3066 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3067
3068 // Make sure we haven't already deoptimized from this tactic.
3069 if (too_many_traps_or_recompiles(reason))
3070 return nullptr;
3071
3072 // (No, this isn't a call, but it's enough like a virtual call
3073 // to use the same ciMethod accessor to get the profile info...)
3074 // If we have a speculative type use it instead of profiling (which
3075 // may not help us)
3076 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3077 if (exact_kls != nullptr) {// no cast failures here
3078 if (require_klass == nullptr ||
3079 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3080 // If we narrow the type to match what the type profile sees or
3081 // the speculative type, we can then remove the rest of the
3082 // cast.
3083 // This is a win, even if the exact_kls is very specific,
3084 // because downstream operations, such as method calls,
3085 // will often benefit from the sharper type.
3086 Node* exact_obj = not_null_obj; // will get updated in place...
3087 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3088 &exact_obj);
3089 { PreserveJVMState pjvms(this);
3090 set_control(slow_ctl);
3091 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3092 }
3093 if (safe_for_replace) {
3094 replace_in_map(not_null_obj, exact_obj);
3095 }
3096 return exact_obj;
3186 // If not_null_obj is dead, only null-path is taken
3187 if (stopped()) { // Doing instance-of on a null?
3188 set_control(null_ctl);
3189 return intcon(0);
3190 }
3191 region->init_req(_null_path, null_ctl);
3192 phi ->init_req(_null_path, intcon(0)); // Set null path value
3193 if (null_ctl == top()) {
3194 // Do this eagerly, so that pattern matches like is_diamond_phi
3195 // will work even during parsing.
3196 assert(_null_path == PATH_LIMIT-1, "delete last");
3197 region->del_req(_null_path);
3198 phi ->del_req(_null_path);
3199 }
3200
3201 // Do we know the type check always succeed?
3202 bool known_statically = false;
3203 if (_gvn.type(superklass)->singleton()) {
3204 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3205 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3206 if (subk->is_loaded()) {
3207 int static_res = C->static_subtype_check(superk, subk);
3208 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3209 }
3210 }
3211
3212 if (!known_statically) {
3213 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3214 // We may not have profiling here or it may not help us. If we
3215 // have a speculative type use it to perform an exact cast.
3216 ciKlass* spec_obj_type = obj_type->speculative_type();
3217 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3218 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3219 if (stopped()) { // Profile disagrees with this path.
3220 set_control(null_ctl); // Null is the only remaining possibility.
3221 return intcon(0);
3222 }
3223 if (cast_obj != nullptr) {
3224 not_null_obj = cast_obj;
3225 }
3226 }
3242 record_for_igvn(region);
3243
3244 // If we know the type check always succeeds then we don't use the
3245 // profiling data at this bytecode. Don't lose it, feed it to the
3246 // type system as a speculative type.
3247 if (safe_for_replace) {
3248 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3249 replace_in_map(obj, casted_obj);
3250 }
3251
3252 return _gvn.transform(phi);
3253 }
3254
3255 //-------------------------------gen_checkcast---------------------------------
3256 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3257 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3258 // uncommon-trap paths work. Adjust stack after this call.
3259 // If failure_control is supplied and not null, it is filled in with
3260 // the control edge for the cast failure. Otherwise, an appropriate
3261 // uncommon trap or exception is thrown.
3262 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3263 Node* *failure_control) {
3264 kill_dead_locals(); // Benefit all the uncommon traps
3265 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3266 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3267 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3268
3269 // Fast cutout: Check the case that the cast is vacuously true.
3270 // This detects the common cases where the test will short-circuit
3271 // away completely. We do this before we perform the null check,
3272 // because if the test is going to turn into zero code, we don't
3273 // want a residual null check left around. (Causes a slowdown,
3274 // for example, in some objArray manipulations, such as a[i]=a[j].)
3275 if (improved_klass_ptr_type->singleton()) {
3276 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3277 if (objtp != nullptr) {
3278 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3279 case Compile::SSC_always_true:
3280 // If we know the type check always succeed then we don't use
3281 // the profiling data at this bytecode. Don't lose it, feed it
3282 // to the type system as a speculative type.
3283 return record_profiled_receiver_for_speculation(obj);
3284 case Compile::SSC_always_false:
3285 // It needs a null check because a null will *pass* the cast check.
3286 // A non-null value will always produce an exception.
3287 if (!objtp->maybe_null()) {
3288 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3289 Deoptimization::DeoptReason reason = is_aastore ?
3290 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3291 builtin_throw(reason);
3292 return top();
3293 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3294 return null_assert(obj);
3295 }
3296 break; // Fall through to full check
3297 default:
3298 break;
3299 }
3300 }
3301 }
3302
3303 ciProfileData* data = nullptr;
3304 bool safe_for_replace = false;
3305 if (failure_control == nullptr) { // use MDO in regular case only
3306 assert(java_bc() == Bytecodes::_aastore ||
3307 java_bc() == Bytecodes::_checkcast,
3308 "interpreter profiles type checks only for these BCs");
3309 data = method()->method_data()->bci_to_data(bci());
3310 safe_for_replace = true;
3311 }
3312
3313 // Make the merge point
3314 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3315 RegionNode* region = new RegionNode(PATH_LIMIT);
3316 Node* phi = new PhiNode(region, toop);
3317 C->set_has_split_ifs(true); // Has chance for split-if optimization
3318
3319 // Use null-cast information if it is available
3320 bool speculative_not_null = false;
3321 bool never_see_null = ((failure_control == nullptr) // regular case only
3322 && seems_never_null(obj, data, speculative_not_null));
3323
3324 // Null check; get casted pointer; set region slot 3
3325 Node* null_ctl = top();
3326 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3327
3328 // If not_null_obj is dead, only null-path is taken
3329 if (stopped()) { // Doing instance-of on a null?
3330 set_control(null_ctl);
3331 return null();
3332 }
3333 region->init_req(_null_path, null_ctl);
3334 phi ->init_req(_null_path, null()); // Set null path value
3335 if (null_ctl == top()) {
3336 // Do this eagerly, so that pattern matches like is_diamond_phi
3337 // will work even during parsing.
3338 assert(_null_path == PATH_LIMIT-1, "delete last");
3339 region->del_req(_null_path);
3340 phi ->del_req(_null_path);
3341 }
3342
3343 Node* cast_obj = nullptr;
3344 if (improved_klass_ptr_type->klass_is_exact()) {
3345 // The following optimization tries to statically cast the speculative type of the object
3346 // (for example obtained during profiling) to the type of the superklass and then do a
3347 // dynamic check that the type of the object is what we expect. To work correctly
3348 // for checkcast and aastore the type of superklass should be exact.
3349 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3350 // We may not have profiling here or it may not help us. If we have
3351 // a speculative type use it to perform an exact cast.
3352 ciKlass* spec_obj_type = obj_type->speculative_type();
3353 if (spec_obj_type != nullptr || data != nullptr) {
3354 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3355 if (cast_obj != nullptr) {
3356 if (failure_control != nullptr) // failure is now impossible
3357 (*failure_control) = top();
3358 // adjust the type of the phi to the exact klass:
3359 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3360 }
3361 }
3362 }
3363
3364 if (cast_obj == nullptr) {
3365 // Generate the subtype check
3366 Node* improved_superklass = superklass;
3367 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3368 improved_superklass = makecon(improved_klass_ptr_type);
3369 }
3370 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3371
3372 // Plug in success path into the merge
3373 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3374 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3375 if (failure_control == nullptr) {
3376 if (not_subtype_ctrl != top()) { // If failure is possible
3377 PreserveJVMState pjvms(this);
3378 set_control(not_subtype_ctrl);
3379 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3380 Deoptimization::DeoptReason reason = is_aastore ?
3381 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3382 builtin_throw(reason);
3383 }
3384 } else {
3385 (*failure_control) = not_subtype_ctrl;
3386 }
3387 }
3388
3389 region->init_req(_obj_path, control());
3390 phi ->init_req(_obj_path, cast_obj);
3391
3392 // A merge of null or Casted-NotNull obj
3393 Node* res = _gvn.transform(phi);
3394
3395 // Note I do NOT always 'replace_in_map(obj,result)' here.
3396 // if( tk->klass()->can_be_primary_super() )
3397 // This means that if I successfully store an Object into an array-of-String
3398 // I 'forget' that the Object is really now known to be a String. I have to
3399 // do this because we don't have true union types for interfaces - if I store
3400 // a Baz into an array-of-Interface and then tell the optimizer it's an
3401 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3402 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3403 // replace_in_map( obj, res );
3404
3405 // Return final merged results
3406 set_control( _gvn.transform(region) );
3407 record_for_igvn(region);
3408
3409 return record_profiled_receiver_for_speculation(res);
3410 }
3411
3412 //------------------------------next_monitor-----------------------------------
3413 // What number should be given to the next monitor?
3414 int GraphKit::next_monitor() {
3415 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3416 int next = current + C->sync_stack_slots();
3417 // Keep the toplevel high water mark current:
3418 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3419 return current;
3420 }
3421
3422 //------------------------------insert_mem_bar---------------------------------
3423 // Memory barrier to avoid floating things around
3424 // The membar serves as a pinch point between both control and all memory slices.
3425 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3426 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3427 mb->init_req(TypeFunc::Control, control());
3428 mb->init_req(TypeFunc::Memory, reset_memory());
3429 Node* membar = _gvn.transform(mb);
3523 lock->create_lock_counter(map()->jvms());
3524 increment_counter(lock->counter()->addr());
3525 }
3526 #endif
3527
3528 return flock;
3529 }
3530
3531
3532 //------------------------------shared_unlock----------------------------------
3533 // Emit unlocking code.
3534 void GraphKit::shared_unlock(Node* box, Node* obj) {
3535 // bci is either a monitorenter bc or InvocationEntryBci
3536 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3537 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3538
3539 if (stopped()) { // Dead monitor?
3540 map()->pop_monitor(); // Kill monitor from debug info
3541 return;
3542 }
3543
3544 // Memory barrier to avoid floating things down past the locked region
3545 insert_mem_bar(Op_MemBarReleaseLock);
3546
3547 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3548 UnlockNode *unlock = new UnlockNode(C, tf);
3549 #ifdef ASSERT
3550 unlock->set_dbg_jvms(sync_jvms());
3551 #endif
3552 uint raw_idx = Compile::AliasIdxRaw;
3553 unlock->init_req( TypeFunc::Control, control() );
3554 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3555 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3556 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3557 unlock->init_req( TypeFunc::ReturnAdr, top() );
3558
3559 unlock->init_req(TypeFunc::Parms + 0, obj);
3560 unlock->init_req(TypeFunc::Parms + 1, box);
3561 unlock = _gvn.transform(unlock)->as_Unlock();
3562
3563 Node* mem = reset_memory();
3564
3565 // unlock has no side-effects, sets few values
3566 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3567
3568 // Kill monitor from debug info
3569 map()->pop_monitor( );
3570 }
3571
3572 //-------------------------------get_layout_helper-----------------------------
3573 // If the given klass is a constant or known to be an array,
3574 // fetch the constant layout helper value into constant_value
3575 // and return null. Otherwise, load the non-constant
3576 // layout helper value, and return the node which represents it.
3577 // This two-faced routine is useful because allocation sites
3578 // almost always feature constant types.
3579 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3580 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3581 if (!StressReflectiveCode && klass_t != nullptr) {
3582 bool xklass = klass_t->klass_is_exact();
3583 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3584 jint lhelper;
3585 if (klass_t->isa_aryklassptr()) {
3586 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3587 if (is_reference_type(elem, true)) {
3588 elem = T_OBJECT;
3589 }
3590 lhelper = Klass::array_layout_helper(elem);
3591 } else {
3592 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3593 }
3594 if (lhelper != Klass::_lh_neutral_value) {
3595 constant_value = lhelper;
3596 return (Node*) nullptr;
3597 }
3598 }
3599 }
3600 constant_value = Klass::_lh_neutral_value; // put in a known value
3601 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3602 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3603 }
3604
3605 // We just put in an allocate/initialize with a big raw-memory effect.
3606 // Hook selected additional alias categories on the initialization.
3607 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3608 MergeMemNode* init_in_merge,
3609 Node* init_out_raw) {
3610 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3611 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3612
3613 Node* prevmem = kit.memory(alias_idx);
3614 init_in_merge->set_memory_at(alias_idx, prevmem);
3615 kit.set_memory(init_out_raw, alias_idx);
3616 }
3617
3618 //---------------------------set_output_for_allocation-------------------------
3619 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3620 const TypeOopPtr* oop_type,
3621 bool deoptimize_on_exception) {
3622 int rawidx = Compile::AliasIdxRaw;
3623 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3624 add_safepoint_edges(alloc);
3625 Node* allocx = _gvn.transform(alloc);
3626 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3627 // create memory projection for i_o
3628 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3629 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3630
3631 // create a memory projection as for the normal control path
3632 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3633 set_memory(malloc, rawidx);
3634
3635 // a normal slow-call doesn't change i_o, but an allocation does
3636 // we create a separate i_o projection for the normal control path
3637 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3638 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3639
3640 // put in an initialization barrier
3641 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3642 rawoop)->as_Initialize();
3643 assert(alloc->initialization() == init, "2-way macro link must work");
3644 assert(init ->allocation() == alloc, "2-way macro link must work");
3645 {
3646 // Extract memory strands which may participate in the new object's
3647 // initialization, and source them from the new InitializeNode.
3648 // This will allow us to observe initializations when they occur,
3649 // and link them properly (as a group) to the InitializeNode.
3650 assert(init->in(InitializeNode::Memory) == malloc, "");
3651 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3652 init->set_req(InitializeNode::Memory, minit_in);
3653 record_for_igvn(minit_in); // fold it up later, if possible
3654 Node* minit_out = memory(rawidx);
3655 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3656 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
3657 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
3658 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
3659 // multiple projections as a result.
3660 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
3661 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
3662 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
3663 if (oop_type->isa_aryptr()) {
3664 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3665 int elemidx = C->get_alias_index(telemref);
3666 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(elemidx))));
3667 } else if (oop_type->isa_instptr()) {
3668 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3669 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3670 ciField* field = ik->nonstatic_field_at(i);
3671 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3672 continue; // do not bother to track really large numbers of fields
3673 // Find (or create) the alias category for this field:
3674 int fieldidx = C->alias_type(field)->index();
3675 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
3676 }
3677 }
3678 }
3679
3680 // Cast raw oop to the real thing...
3681 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3682 javaoop = _gvn.transform(javaoop);
3683 C->set_recent_alloc(control(), javaoop);
3684 assert(just_allocated_object(control()) == javaoop, "just allocated");
3685
3686 #ifdef ASSERT
3698 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3699 }
3700 }
3701 #endif //ASSERT
3702
3703 return javaoop;
3704 }
3705
3706 //---------------------------new_instance--------------------------------------
3707 // This routine takes a klass_node which may be constant (for a static type)
3708 // or may be non-constant (for reflective code). It will work equally well
3709 // for either, and the graph will fold nicely if the optimizer later reduces
3710 // the type to a constant.
3711 // The optional arguments are for specialized use by intrinsics:
3712 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3713 // - If 'return_size_val', report the total object size to the caller.
3714 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3715 Node* GraphKit::new_instance(Node* klass_node,
3716 Node* extra_slow_test,
3717 Node* *return_size_val,
3718 bool deoptimize_on_exception) {
3719 // Compute size in doublewords
3720 // The size is always an integral number of doublewords, represented
3721 // as a positive bytewise size stored in the klass's layout_helper.
3722 // The layout_helper also encodes (in a low bit) the need for a slow path.
3723 jint layout_con = Klass::_lh_neutral_value;
3724 Node* layout_val = get_layout_helper(klass_node, layout_con);
3725 int layout_is_con = (layout_val == nullptr);
3726
3727 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3728 // Generate the initial go-slow test. It's either ALWAYS (return a
3729 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3730 // case) a computed value derived from the layout_helper.
3731 Node* initial_slow_test = nullptr;
3732 if (layout_is_con) {
3733 assert(!StressReflectiveCode, "stress mode does not use these paths");
3734 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3735 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3736 } else { // reflective case
3737 // This reflective path is used by Unsafe.allocateInstance.
3738 // (It may be stress-tested by specifying StressReflectiveCode.)
3739 // Basically, we want to get into the VM is there's an illegal argument.
3740 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3741 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3742 if (extra_slow_test != intcon(0)) {
3743 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3744 }
3745 // (Macro-expander will further convert this to a Bool, if necessary.)
3756
3757 // Clear the low bits to extract layout_helper_size_in_bytes:
3758 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3759 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3760 size = _gvn.transform( new AndXNode(size, mask) );
3761 }
3762 if (return_size_val != nullptr) {
3763 (*return_size_val) = size;
3764 }
3765
3766 // This is a precise notnull oop of the klass.
3767 // (Actually, it need not be precise if this is a reflective allocation.)
3768 // It's what we cast the result to.
3769 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3770 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3771 const TypeOopPtr* oop_type = tklass->as_instance_type();
3772
3773 // Now generate allocation code
3774
3775 // The entire memory state is needed for slow path of the allocation
3776 // since GC and deoptimization can happened.
3777 Node *mem = reset_memory();
3778 set_all_memory(mem); // Create new memory state
3779
3780 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3781 control(), mem, i_o(),
3782 size, klass_node,
3783 initial_slow_test);
3784
3785 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3786 }
3787
3788 //-------------------------------new_array-------------------------------------
3789 // helper for both newarray and anewarray
3790 // The 'length' parameter is (obviously) the length of the array.
3791 // The optional arguments are for specialized use by intrinsics:
3792 // - If 'return_size_val', report the non-padded array size (sum of header size
3793 // and array body) to the caller.
3794 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3795 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3796 Node* length, // number of array elements
3797 int nargs, // number of arguments to push back for uncommon trap
3798 Node* *return_size_val,
3799 bool deoptimize_on_exception) {
3800 jint layout_con = Klass::_lh_neutral_value;
3801 Node* layout_val = get_layout_helper(klass_node, layout_con);
3802 int layout_is_con = (layout_val == nullptr);
3803
3804 if (!layout_is_con && !StressReflectiveCode &&
3805 !too_many_traps(Deoptimization::Reason_class_check)) {
3806 // This is a reflective array creation site.
3807 // Optimistically assume that it is a subtype of Object[],
3808 // so that we can fold up all the address arithmetic.
3809 layout_con = Klass::array_layout_helper(T_OBJECT);
3810 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3811 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3812 { BuildCutout unless(this, bol_lh, PROB_MAX);
3813 inc_sp(nargs);
3814 uncommon_trap(Deoptimization::Reason_class_check,
3815 Deoptimization::Action_maybe_recompile);
3816 }
3817 layout_val = nullptr;
3818 layout_is_con = true;
3819 }
3820
3821 // Generate the initial go-slow test. Make sure we do not overflow
3822 // if length is huge (near 2Gig) or negative! We do not need
3823 // exact double-words here, just a close approximation of needed
3824 // double-words. We can't add any offset or rounding bits, lest we
3825 // take a size -1 of bytes and make it positive. Use an unsigned
3826 // compare, so negative sizes look hugely positive.
3827 int fast_size_limit = FastAllocateSizeLimit;
3828 if (layout_is_con) {
3829 assert(!StressReflectiveCode, "stress mode does not use these paths");
3830 // Increase the size limit if we have exact knowledge of array type.
3831 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3832 assert(fast_size_limit == 0 || count_leading_zeros(fast_size_limit) > static_cast<unsigned>(LogBytesPerLong - log2_esize),
3833 "fast_size_limit (%d) overflow when shifted left by %d", fast_size_limit, LogBytesPerLong - log2_esize);
3834 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3835 }
3836
3837 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3838 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3839
3840 // --- Size Computation ---
3841 // array_size = round_to_heap(array_header + (length << elem_shift));
3842 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3843 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3844 // The rounding mask is strength-reduced, if possible.
3845 int round_mask = MinObjAlignmentInBytes - 1;
3846 Node* header_size = nullptr;
3847 // (T_BYTE has the weakest alignment and size restrictions...)
3848 if (layout_is_con) {
3849 int hsize = Klass::layout_helper_header_size(layout_con);
3850 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3851 if ((round_mask & ~right_n_bits(eshift)) == 0)
3852 round_mask = 0; // strength-reduce it if it goes away completely
3853 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3854 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3855 assert(header_size_min <= hsize, "generic minimum is smallest");
3856 header_size = intcon(hsize);
3857 } else {
3858 Node* hss = intcon(Klass::_lh_header_size_shift);
3859 Node* hsm = intcon(Klass::_lh_header_size_mask);
3860 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3861 header_size = _gvn.transform(new AndINode(header_size, hsm));
3862 }
3863
3864 Node* elem_shift = nullptr;
3865 if (layout_is_con) {
3866 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3867 if (eshift != 0)
3868 elem_shift = intcon(eshift);
3869 } else {
3870 // There is no need to mask or shift this value.
3871 // The semantics of LShiftINode include an implicit mask to 0x1F.
3872 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3873 elem_shift = layout_val;
3922 }
3923 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
3924
3925 if (return_size_val != nullptr) {
3926 // This is the size
3927 (*return_size_val) = non_rounded_size;
3928 }
3929
3930 Node* size = non_rounded_size;
3931 if (round_mask != 0) {
3932 Node* mask1 = MakeConX(round_mask);
3933 size = _gvn.transform(new AddXNode(size, mask1));
3934 Node* mask2 = MakeConX(~round_mask);
3935 size = _gvn.transform(new AndXNode(size, mask2));
3936 }
3937 // else if round_mask == 0, the size computation is self-rounding
3938
3939 // Now generate allocation code
3940
3941 // The entire memory state is needed for slow path of the allocation
3942 // since GC and deoptimization can happened.
3943 Node *mem = reset_memory();
3944 set_all_memory(mem); // Create new memory state
3945
3946 if (initial_slow_test->is_Bool()) {
3947 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3948 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3949 }
3950
3951 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3952 Node* valid_length_test = _gvn.intcon(1);
3953 if (ary_type->isa_aryptr()) {
3954 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
3955 jint max = TypeAryPtr::max_array_length(bt);
3956 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
3957 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
3958 }
3959
3960 // Create the AllocateArrayNode and its result projections
3961 AllocateArrayNode* alloc
3962 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3963 control(), mem, i_o(),
3964 size, klass_node,
3965 initial_slow_test,
3966 length, valid_length_test);
3967
3968 // Cast to correct type. Note that the klass_node may be constant or not,
3969 // and in the latter case the actual array type will be inexact also.
3970 // (This happens via a non-constant argument to inline_native_newArray.)
3971 // In any case, the value of klass_node provides the desired array type.
3972 const TypeInt* length_type = _gvn.find_int_type(length);
3973 if (ary_type->isa_aryptr() && length_type != nullptr) {
3974 // Try to get a better type than POS for the size
3975 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3976 }
3977
3978 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3979
3980 array_ideal_length(alloc, ary_type, true);
3981 return javaoop;
3982 }
3983
3984 // The following "Ideal_foo" functions are placed here because they recognize
3985 // the graph shapes created by the functions immediately above.
3986
3987 //---------------------------Ideal_allocation----------------------------------
4082 void GraphKit::add_parse_predicates(int nargs) {
4083 if (ShortRunningLongLoop) {
4084 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4085 // walking up from the loop.
4086 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4087 }
4088 if (UseLoopPredicate) {
4089 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4090 if (UseProfiledLoopPredicate) {
4091 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4092 }
4093 }
4094 if (UseAutoVectorizationPredicate) {
4095 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4096 }
4097 // Loop Limit Check Predicate should be near the loop.
4098 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4099 }
4100
4101 void GraphKit::sync_kit(IdealKit& ideal) {
4102 set_all_memory(ideal.merged_memory());
4103 set_i_o(ideal.i_o());
4104 set_control(ideal.ctrl());
4105 }
4106
4107 void GraphKit::final_sync(IdealKit& ideal) {
4108 // Final sync IdealKit and graphKit.
4109 sync_kit(ideal);
4110 }
4111
4112 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4113 Node* len = load_array_length(load_String_value(str, set_ctrl));
4114 Node* coder = load_String_coder(str, set_ctrl);
4115 // Divide length by 2 if coder is UTF16
4116 return _gvn.transform(new RShiftINode(len, coder));
4117 }
4118
4119 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4120 int value_offset = java_lang_String::value_offset();
4121 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4122 false, nullptr, 0);
4123 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4124 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4125 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4126 ciTypeArrayKlass::make(T_BYTE), true, 0);
4127 Node* p = basic_plus_adr(str, str, value_offset);
4128 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4129 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4130 return load;
4131 }
4132
4133 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4134 if (!CompactStrings) {
4135 return intcon(java_lang_String::CODER_UTF16);
4136 }
4137 int coder_offset = java_lang_String::coder_offset();
4138 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4139 false, nullptr, 0);
4140 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4141
4142 Node* p = basic_plus_adr(str, str, coder_offset);
4143 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4144 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4145 return load;
4146 }
4147
4148 void GraphKit::store_String_value(Node* str, Node* value) {
4149 int value_offset = java_lang_String::value_offset();
4150 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4151 false, nullptr, 0);
4152 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4153
4154 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4155 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4156 }
4157
4158 void GraphKit::store_String_coder(Node* str, Node* value) {
4159 int coder_offset = java_lang_String::coder_offset();
4160 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4161 false, nullptr, 0);
4162 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4163
4164 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4165 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4166 }
4167
4168 // Capture src and dst memory state with a MergeMemNode
4169 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4170 if (src_type == dst_type) {
4171 // Types are equal, we don't need a MergeMemNode
4172 return memory(src_type);
4173 }
4174 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4175 record_for_igvn(merge); // fold it up later, if possible
4176 int src_idx = C->get_alias_index(src_type);
4177 int dst_idx = C->get_alias_index(dst_type);
4178 merge->set_memory_at(src_idx, memory(src_idx));
4179 merge->set_memory_at(dst_idx, memory(dst_idx));
4180 return merge;
4181 }
4254 i_char->init_req(2, AddI(i_char, intcon(2)));
4255
4256 set_control(IfFalse(iff));
4257 set_memory(st, TypeAryPtr::BYTES);
4258 }
4259
4260 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4261 if (!field->is_constant()) {
4262 return nullptr; // Field not marked as constant.
4263 }
4264 ciInstance* holder = nullptr;
4265 if (!field->is_static()) {
4266 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4267 if (const_oop != nullptr && const_oop->is_instance()) {
4268 holder = const_oop->as_instance();
4269 }
4270 }
4271 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4272 /*is_unsigned_load=*/false);
4273 if (con_type != nullptr) {
4274 return makecon(con_type);
4275 }
4276 return nullptr;
4277 }
4278
4279 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4280 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4281 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4282 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4283 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4284 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4285 return casted_obj;
4286 }
4287 return obj;
4288 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciMethod.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "ci/ciUtilities.hpp"
31 #include "classfile/javaClasses.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/flatArrayKlass.hpp"
38 #include "opto/addnode.hpp"
39 #include "opto/castnode.hpp"
40 #include "opto/convertnode.hpp"
41 #include "opto/graphKit.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/inlinetypenode.hpp"
44 #include "opto/intrinsicnode.hpp"
45 #include "opto/locknode.hpp"
46 #include "opto/machnode.hpp"
47 #include "opto/multnode.hpp"
48 #include "opto/narrowptrnode.hpp"
49 #include "opto/opaquenode.hpp"
50 #include "opto/parse.hpp"
51 #include "opto/rootnode.hpp"
52 #include "opto/runtime.hpp"
53 #include "opto/subtypenode.hpp"
54 #include "runtime/arguments.hpp"
55 #include "runtime/deoptimization.hpp"
56 #include "runtime/sharedRuntime.hpp"
57 #include "runtime/stubRoutines.hpp"
58 #include "utilities/bitMap.inline.hpp"
59 #include "utilities/growableArray.hpp"
60 #include "utilities/powerOfTwo.hpp"
61
62 //----------------------------GraphKit-----------------------------------------
63 // Main utility constructor.
64 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
65 : Phase(Phase::Parser),
66 _env(C->env()),
67 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
68 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
69 {
70 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
71 _exceptions = jvms->map()->next_exception();
72 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
73 set_jvms(jvms);
74 #ifdef ASSERT
75 if (_gvn.is_IterGVN() != nullptr) {
76 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
77 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
78 _worklist_size = _gvn.C->igvn_worklist()->size();
79 }
80 #endif
81 }
82
83 // Private constructor for parser.
84 GraphKit::GraphKit()
85 : Phase(Phase::Parser),
86 _env(C->env()),
87 _gvn(*C->initial_gvn()),
88 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
89 {
90 _exceptions = nullptr;
91 set_map(nullptr);
92 DEBUG_ONLY(_sp = -99);
93 DEBUG_ONLY(set_bci(-99));
94 }
95
96
97
98 //---------------------------clean_stack---------------------------------------
99 // Clear away rubbish from the stack area of the JVM state.
100 // This destroys any arguments that may be waiting on the stack.
345 }
346 static inline void add_one_req(Node* dstphi, Node* src) {
347 assert(is_hidden_merge(dstphi), "must be a special merge node");
348 assert(!is_hidden_merge(src), "must not be a special merge node");
349 dstphi->add_req(src);
350 }
351
352 //-----------------------combine_exception_states------------------------------
353 // This helper function combines exception states by building phis on a
354 // specially marked state-merging region. These regions and phis are
355 // untransformed, and can build up gradually. The region is marked by
356 // having a control input of its exception map, rather than null. Such
357 // regions do not appear except in this function, and in use_exception_state.
358 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
359 if (failing_internal()) {
360 return; // dying anyway...
361 }
362 JVMState* ex_jvms = ex_map->_jvms;
363 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
364 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
365 // TODO 8325632 Re-enable
366 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
367 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
368 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
369 assert(ex_map->req() == phi_map->req(), "matching maps");
370 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
371 Node* hidden_merge_mark = root();
372 Node* region = phi_map->control();
373 MergeMemNode* phi_mem = phi_map->merged_memory();
374 MergeMemNode* ex_mem = ex_map->merged_memory();
375 if (region->in(0) != hidden_merge_mark) {
376 // The control input is not (yet) a specially-marked region in phi_map.
377 // Make it so, and build some phis.
378 region = new RegionNode(2);
379 _gvn.set_type(region, Type::CONTROL);
380 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
381 region->init_req(1, phi_map->control());
382 phi_map->set_control(region);
383 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
384 record_for_igvn(io_phi);
385 _gvn.set_type(io_phi, Type::ABIO);
386 phi_map->set_i_o(io_phi);
874 if (PrintMiscellaneous && (Verbose || WizardMode)) {
875 tty->print_cr("Zombie local %d: ", local);
876 jvms->dump();
877 }
878 return false;
879 }
880 }
881 }
882 return true;
883 }
884
885 #endif //ASSERT
886
887 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
888 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
889 ciMethod* cur_method = jvms->method();
890 int cur_bci = jvms->bci();
891 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
892 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
893 return Interpreter::bytecode_should_reexecute(code) ||
894 (is_anewarray && (code == Bytecodes::_multianewarray));
895 // Reexecute _multianewarray bytecode which was replaced with
896 // sequence of [a]newarray. See Parse::do_multianewarray().
897 //
898 // Note: interpreter should not have it set since this optimization
899 // is limited by dimensions and guarded by flag so in some cases
900 // multianewarray() runtime calls will be generated and
901 // the bytecode should not be reexecutes (stack will not be reset).
902 } else {
903 return false;
904 }
905 }
906
907 // Helper function for adding JVMState and debug information to node
908 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
909 // Add the safepoint edges to the call (or other safepoint).
910
911 // Make sure dead locals are set to top. This
912 // should help register allocation time and cut down on the size
913 // of the deoptimization information.
914 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
942
943 if (env()->should_retain_local_variables()) {
944 // At any safepoint, this method can get breakpointed, which would
945 // then require an immediate deoptimization.
946 can_prune_locals = false; // do not prune locals
947 stack_slots_not_pruned = 0;
948 }
949
950 // do not scribble on the input jvms
951 JVMState* out_jvms = youngest_jvms->clone_deep(C);
952 call->set_jvms(out_jvms); // Start jvms list for call node
953
954 // For a known set of bytecodes, the interpreter should reexecute them if
955 // deoptimization happens. We set the reexecute state for them here
956 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
957 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
958 #ifdef ASSERT
959 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
960 assert(method() == youngest_jvms->method(), "sanity");
961 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
962 // TODO 8371125
963 // assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
964 #endif // ASSERT
965 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
966 }
967
968 // Presize the call:
969 DEBUG_ONLY(uint non_debug_edges = call->req());
970 call->add_req_batch(top(), youngest_jvms->debug_depth());
971 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
972
973 // Set up edges so that the call looks like this:
974 // Call [state:] ctl io mem fptr retadr
975 // [parms:] parm0 ... parmN
976 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
977 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
978 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
979 // Note that caller debug info precedes callee debug info.
980
981 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
982 uint debug_ptr = call->req();
983
984 // Loop over the map input edges associated with jvms, add them
985 // to the call node, & reset all offsets to match call node array.
986
987 JVMState* callee_jvms = nullptr;
988 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
989 uint debug_end = debug_ptr;
990 uint debug_start = debug_ptr - in_jvms->debug_size();
991 debug_ptr = debug_start; // back up the ptr
992
993 uint p = debug_start; // walks forward in [debug_start, debug_end)
994 uint j, k, l;
995 SafePointNode* in_map = in_jvms->map();
996 out_jvms->set_map(call);
997
998 if (can_prune_locals) {
999 assert(in_jvms->method() == out_jvms->method(), "sanity");
1000 // If the current throw can reach an exception handler in this JVMS,
1001 // then we must keep everything live that can reach that handler.
1002 // As a quick and dirty approximation, we look for any handlers at all.
1003 if (in_jvms->method()->has_exception_handlers()) {
1004 can_prune_locals = false;
1005 }
1006 }
1007
1008 // Add the Locals
1009 k = in_jvms->locoff();
1010 l = in_jvms->loc_size();
1011 out_jvms->set_locoff(p);
1012 if (!can_prune_locals) {
1013 for (j = 0; j < l; j++) {
1014 call->set_req(p++, in_map->in(k + j));
1015 }
1016 } else {
1017 p += l; // already set to top above by add_req_batch
1018 }
1019
1020 // Add the Expression Stack
1021 k = in_jvms->stkoff();
1022 l = in_jvms->sp();
1023 out_jvms->set_stkoff(p);
1024 if (!can_prune_locals) {
1025 for (j = 0; j < l; j++) {
1026 call->set_req(p++, in_map->in(k + j));
1027 }
1028 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1029 // Divide stack into {S0,...,S1}, where S0 is set to top.
1030 uint s1 = stack_slots_not_pruned;
1031 stack_slots_not_pruned = 0; // for next iteration
1032 if (s1 > l) s1 = l;
1033 uint s0 = l - s1;
1034 p += s0; // skip the tops preinstalled by add_req_batch
1035 for (j = s0; j < l; j++)
1036 call->set_req(p++, in_map->in(k+j));
1037 } else {
1038 p += l; // already set to top above by add_req_batch
1039 }
1040
1041 // Add the Monitors
1042 k = in_jvms->monoff();
1043 l = in_jvms->mon_size();
1044 out_jvms->set_monoff(p);
1045 for (j = 0; j < l; j++)
1046 call->set_req(p++, in_map->in(k+j));
1047
1048 // Copy any scalar object fields.
1049 k = in_jvms->scloff();
1050 l = in_jvms->scl_size();
1051 out_jvms->set_scloff(p);
1052 for (j = 0; j < l; j++)
1053 call->set_req(p++, in_map->in(k+j));
1054
1055 // Finish the new jvms.
1056 out_jvms->set_endoff(p);
1057
1058 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1059 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1060 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1061 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1062 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1063 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1064
1065 // Update the two tail pointers in parallel.
1066 callee_jvms = out_jvms;
1067 out_jvms = out_jvms->caller();
1068 in_jvms = in_jvms->caller();
1069 }
1070
1071 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1072
1073 // Test the correctness of JVMState::debug_xxx accessors:
1074 assert(call->jvms()->debug_start() == non_debug_edges, "");
1075 assert(call->jvms()->debug_end() == call->req(), "");
1076 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1077 }
1078
1079 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1080 Bytecodes::Code code = java_bc();
1081 if (code == Bytecodes::_wide) {
1082 code = method()->java_code_at_bci(bci() + 1);
1083 }
1084
1085 if (code != Bytecodes::_illegal) {
1086 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1222 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1223 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1224 return _gvn.transform( new AndLNode(conv, mask) );
1225 }
1226
1227 Node* GraphKit::ConvL2I(Node* offset) {
1228 // short-circuit a common case
1229 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1230 if (offset_con != (jlong)Type::OffsetBot) {
1231 return intcon((int) offset_con);
1232 }
1233 return _gvn.transform( new ConvL2INode(offset));
1234 }
1235
1236 //-------------------------load_object_klass-----------------------------------
1237 Node* GraphKit::load_object_klass(Node* obj) {
1238 // Special-case a fresh allocation to avoid building nodes:
1239 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1240 if (akls != nullptr) return akls;
1241 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1242 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1243 }
1244
1245 //-------------------------load_array_length-----------------------------------
1246 Node* GraphKit::load_array_length(Node* array) {
1247 // Special-case a fresh allocation to avoid building nodes:
1248 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1249 Node *alen;
1250 if (alloc == nullptr) {
1251 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1252 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1253 } else {
1254 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1255 }
1256 return alen;
1257 }
1258
1259 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1260 const TypeOopPtr* oop_type,
1261 bool replace_length_in_map) {
1262 Node* length = alloc->Ideal_length();
1271 replace_in_map(length, ccast);
1272 }
1273 return ccast;
1274 }
1275 }
1276 return length;
1277 }
1278
1279 //------------------------------do_null_check----------------------------------
1280 // Helper function to do a null pointer check. Returned value is
1281 // the incoming address with null casted away. You are allowed to use the
1282 // not-null value only if you are control dependent on the test.
1283 #ifndef PRODUCT
1284 extern uint explicit_null_checks_inserted,
1285 explicit_null_checks_elided;
1286 #endif
1287 Node* GraphKit::null_check_common(Node* value, BasicType type,
1288 // optional arguments for variations:
1289 bool assert_null,
1290 Node* *null_control,
1291 bool speculative,
1292 bool null_marker_check) {
1293 assert(!assert_null || null_control == nullptr, "not both at once");
1294 if (stopped()) return top();
1295 NOT_PRODUCT(explicit_null_checks_inserted++);
1296
1297 if (value->is_InlineType()) {
1298 // Null checking a scalarized but nullable inline type. Check the null marker
1299 // input instead of the oop input to avoid keeping buffer allocations alive.
1300 InlineTypeNode* vtptr = value->as_InlineType();
1301 while (vtptr->get_oop()->is_InlineType()) {
1302 vtptr = vtptr->get_oop()->as_InlineType();
1303 }
1304 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1305 if (stopped()) {
1306 return top();
1307 }
1308 if (assert_null) {
1309 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1310 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1311 // replace_in_map(value, vtptr);
1312 // return vtptr;
1313 replace_in_map(value, null());
1314 return null();
1315 }
1316 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1317 return cast_not_null(value, do_replace_in_map);
1318 }
1319
1320 // Construct null check
1321 Node *chk = nullptr;
1322 switch(type) {
1323 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1324 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1325 case T_ARRAY : // fall through
1326 type = T_OBJECT; // simplify further tests
1327 case T_OBJECT : {
1328 const Type *t = _gvn.type( value );
1329
1330 const TypeOopPtr* tp = t->isa_oopptr();
1331 if (tp != nullptr && !tp->is_loaded()
1332 // Only for do_null_check, not any of its siblings:
1333 && !assert_null && null_control == nullptr) {
1334 // Usually, any field access or invocation on an unloaded oop type
1335 // will simply fail to link, since the statically linked class is
1336 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1337 // the static class is loaded but the sharper oop type is not.
1338 // Rather than checking for this obscure case in lots of places,
1339 // we simply observe that a null check on an unloaded class
1403 }
1404 Node *oldcontrol = control();
1405 set_control(cfg);
1406 Node *res = cast_not_null(value);
1407 set_control(oldcontrol);
1408 NOT_PRODUCT(explicit_null_checks_elided++);
1409 return res;
1410 }
1411 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1412 if (cfg == nullptr) break; // Quit at region nodes
1413 depth++;
1414 }
1415 }
1416
1417 //-----------
1418 // Branch to failure if null
1419 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1420 Deoptimization::DeoptReason reason;
1421 if (assert_null) {
1422 reason = Deoptimization::reason_null_assert(speculative);
1423 } else if (type == T_OBJECT || null_marker_check) {
1424 reason = Deoptimization::reason_null_check(speculative);
1425 } else {
1426 reason = Deoptimization::Reason_div0_check;
1427 }
1428 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1429 // ciMethodData::has_trap_at will return a conservative -1 if any
1430 // must-be-null assertion has failed. This could cause performance
1431 // problems for a method after its first do_null_assert failure.
1432 // Consider using 'Reason_class_check' instead?
1433
1434 // To cause an implicit null check, we set the not-null probability
1435 // to the maximum (PROB_MAX). For an explicit check the probability
1436 // is set to a smaller value.
1437 if (null_control != nullptr || too_many_traps(reason)) {
1438 // probability is less likely
1439 ok_prob = PROB_LIKELY_MAG(3);
1440 } else if (!assert_null &&
1441 (ImplicitNullCheckThreshold > 0) &&
1442 method() != nullptr &&
1443 (method()->method_data()->trap_count(reason)
1477 }
1478
1479 if (assert_null) {
1480 // Cast obj to null on this path.
1481 replace_in_map(value, zerocon(type));
1482 return zerocon(type);
1483 }
1484
1485 // Cast obj to not-null on this path, if there is no null_control.
1486 // (If there is a null_control, a non-null value may come back to haunt us.)
1487 if (type == T_OBJECT) {
1488 Node* cast = cast_not_null(value, false);
1489 if (null_control == nullptr || (*null_control) == top())
1490 replace_in_map(value, cast);
1491 value = cast;
1492 }
1493
1494 return value;
1495 }
1496
1497 //------------------------------cast_not_null----------------------------------
1498 // Cast obj to not-null on this path
1499 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1500 if (obj->is_InlineType()) {
1501 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1502 vt->as_InlineType()->set_null_marker(_gvn);
1503 vt = _gvn.transform(vt);
1504 if (do_replace_in_map) {
1505 replace_in_map(obj, vt);
1506 }
1507 return vt;
1508 }
1509 const Type *t = _gvn.type(obj);
1510 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1511 // Object is already not-null?
1512 if( t == t_not_null ) return obj;
1513
1514 Node* cast = new CastPPNode(control(), obj,t_not_null);
1515 cast = _gvn.transform( cast );
1516
1517 // Scan for instances of 'obj' in the current JVM mapping.
1518 // These instances are known to be not-null after the test.
1519 if (do_replace_in_map)
1520 replace_in_map(obj, cast);
1521
1522 return cast; // Return casted value
1523 }
1524
1525 Node* GraphKit::cast_to_non_larval(Node* obj) {
1526 const Type* obj_type = gvn().type(obj);
1527 if (obj->is_InlineType() || !obj_type->is_inlinetypeptr()) {
1528 return obj;
1529 }
1530
1531 Node* new_obj = InlineTypeNode::make_from_oop(this, obj, obj_type->inline_klass());
1532 replace_in_map(obj, new_obj);
1533 return new_obj;
1534 }
1535
1536 // Sometimes in intrinsics, we implicitly know an object is not null
1537 // (there's no actual null check) so we can cast it to not null. In
1538 // the course of optimizations, the input to the cast can become null.
1539 // In that case that data path will die and we need the control path
1540 // to become dead as well to keep the graph consistent. So we have to
1541 // add a check for null for which one branch can't be taken. It uses
1542 // an OpaqueNotNull node that will cause the check to be removed after loop
1543 // opts so the test goes away and the compiled code doesn't execute a
1544 // useless check.
1545 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1546 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1547 return value;
1548 }
1549 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1550 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1551 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1552 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1553 _gvn.set_type(iff, iff->Value(&_gvn));
1554 if (!tst->is_Con()) {
1555 record_for_igvn(iff);
1627 // These are layered on top of the factory methods in LoadNode and StoreNode,
1628 // and integrate with the parser's memory state and _gvn engine.
1629 //
1630
1631 // factory methods in "int adr_idx"
1632 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1633 MemNode::MemOrd mo,
1634 LoadNode::ControlDependency control_dependency,
1635 bool require_atomic_access,
1636 bool unaligned,
1637 bool mismatched,
1638 bool unsafe,
1639 uint8_t barrier_data) {
1640 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1641 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1642 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1643 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1644 Node* mem = memory(adr_idx);
1645 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1646 ld = _gvn.transform(ld);
1647
1648 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1649 // Improve graph before escape analysis and boxing elimination.
1650 record_for_igvn(ld);
1651 if (ld->is_DecodeN()) {
1652 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1653 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1654 // a Phi). Recording such cases is still perfectly sound, but may be
1655 // unnecessary and result in some minor IGVN overhead.
1656 record_for_igvn(ld->in(1));
1657 }
1658 }
1659 return ld;
1660 }
1661
1662 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1663 MemNode::MemOrd mo,
1664 bool require_atomic_access,
1665 bool unaligned,
1666 bool mismatched,
1667 bool unsafe,
1681 if (unsafe) {
1682 st->as_Store()->set_unsafe_access();
1683 }
1684 st->as_Store()->set_barrier_data(barrier_data);
1685 st = _gvn.transform(st);
1686 set_memory(st, adr_idx);
1687 // Back-to-back stores can only remove intermediate store with DU info
1688 // so push on worklist for optimizer.
1689 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1690 record_for_igvn(st);
1691
1692 return st;
1693 }
1694
1695 Node* GraphKit::access_store_at(Node* obj,
1696 Node* adr,
1697 const TypePtr* adr_type,
1698 Node* val,
1699 const Type* val_type,
1700 BasicType bt,
1701 DecoratorSet decorators,
1702 bool safe_for_replace,
1703 const InlineTypeNode* vt) {
1704 // Transformation of a value which could be null pointer (CastPP #null)
1705 // could be delayed during Parse (for example, in adjust_map_after_if()).
1706 // Execute transformation here to avoid barrier generation in such case.
1707 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1708 val = _gvn.makecon(TypePtr::NULL_PTR);
1709 }
1710
1711 if (stopped()) {
1712 return top(); // Dead path ?
1713 }
1714
1715 assert(val != nullptr, "not dead path");
1716 if (val->is_InlineType()) {
1717 // Store to non-flat field. Buffer the inline type and make sure
1718 // the store is re-executed if the allocation triggers deoptimization.
1719 PreserveReexecuteState preexecs(this);
1720 jvms()->set_should_reexecute(true);
1721 val = val->as_InlineType()->buffer(this, safe_for_replace);
1722 }
1723
1724 C2AccessValuePtr addr(adr, adr_type);
1725 C2AccessValue value(val, val_type);
1726 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1727 if (access.is_raw()) {
1728 return _barrier_set->BarrierSetC2::store_at(access, value);
1729 } else {
1730 return _barrier_set->store_at(access, value);
1731 }
1732 }
1733
1734 Node* GraphKit::access_load_at(Node* obj, // containing obj
1735 Node* adr, // actual address to store val at
1736 const TypePtr* adr_type,
1737 const Type* val_type,
1738 BasicType bt,
1739 DecoratorSet decorators,
1740 Node* ctl) {
1741 if (stopped()) {
1742 return top(); // Dead path ?
1743 }
1744
1745 C2AccessValuePtr addr(adr, adr_type);
1746 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1747 if (access.is_raw()) {
1748 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1749 } else {
1750 return _barrier_set->load_at(access, val_type);
1751 }
1752 }
1753
1754 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1755 const Type* val_type,
1756 BasicType bt,
1757 DecoratorSet decorators) {
1758 if (stopped()) {
1759 return top(); // Dead path ?
1760 }
1761
1762 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1763 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1764 if (access.is_raw()) {
1765 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1766 } else {
1831 Node* new_val,
1832 const Type* value_type,
1833 BasicType bt,
1834 DecoratorSet decorators) {
1835 C2AccessValuePtr addr(adr, adr_type);
1836 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1837 if (access.is_raw()) {
1838 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1839 } else {
1840 return _barrier_set->atomic_add_at(access, new_val, value_type);
1841 }
1842 }
1843
1844 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1845 return _barrier_set->clone(this, src, dst, size, is_array);
1846 }
1847
1848 //-------------------------array_element_address-------------------------
1849 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1850 const TypeInt* sizetype, Node* ctrl) {
1851 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1852 uint shift;
1853 uint header;
1854 if (arytype->is_flat() && arytype->klass_is_exact()) {
1855 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1856 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1857 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1858 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1859 // though we don't need the address node in this case and throw it away again.
1860 shift = arytype->flat_log_elem_size();
1861 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1862 } else {
1863 shift = exact_log2(type2aelembytes(elembt));
1864 header = arrayOopDesc::base_offset_in_bytes(elembt);
1865 }
1866
1867 // short-circuit a common case (saves lots of confusing waste motion)
1868 jint idx_con = find_int_con(idx, -1);
1869 if (idx_con >= 0) {
1870 intptr_t offset = header + ((intptr_t)idx_con << shift);
1871 return basic_plus_adr(ary, offset);
1872 }
1873
1874 // must be correct type for alignment purposes
1875 Node* base = basic_plus_adr(ary, header);
1876 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1877 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1878 return basic_plus_adr(ary, base, scale);
1879 }
1880
1881 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1882 assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1883 if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1884 return cast_to_flat_array_exact(array, elem_vk, true, false);
1885 } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1886 return cast_to_flat_array_exact(array, elem_vk, true, true);
1887 } else if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1888 return cast_to_flat_array_exact(array, elem_vk, false, true);
1889 }
1890
1891 bool is_null_free = false;
1892 if (!elem_vk->has_nullable_atomic_layout()) {
1893 // Element does not have a nullable flat layout, cannot be nullable
1894 is_null_free = true;
1895 }
1896
1897 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1898 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1899 arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1900 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1901 }
1902
1903 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1904 assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1905 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1906 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1907 assert(arytype->klass_is_exact(), "inconsistency");
1908 assert(arytype->is_flat(), "inconsistency");
1909 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1910 assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1911 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1912 }
1913
1914 //-------------------------load_array_element-------------------------
1915 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1916 const Type* elemtype = arytype->elem();
1917 BasicType elembt = elemtype->array_element_basic_type();
1918 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1919 if (elembt == T_NARROWOOP) {
1920 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1921 }
1922 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1923 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1924 return ld;
1925 }
1926
1927 //-------------------------set_arguments_for_java_call-------------------------
1928 // Arguments (pre-popped from the stack) are taken from the JVMS.
1929 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1930 PreserveReexecuteState preexecs(this);
1931 if (Arguments::is_valhalla_enabled()) {
1932 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1933 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1934 jvms()->set_should_reexecute(true);
1935 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1936 inc_sp(arg_size);
1937 }
1938 // Add the call arguments
1939 const TypeTuple* domain = call->tf()->domain_sig();
1940 uint nargs = domain->cnt();
1941 int arg_num = 0;
1942 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1943 uint arg_idx = i - TypeFunc::Parms;
1944 Node* arg = argument(arg_idx);
1945 const Type* t = domain->field_at(i);
1946 // TODO 8284443 A static call to a mismatched method should still be scalarized
1947 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1948 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1949 if (!arg->is_InlineType()) {
1950 // There are 2 cases in which the argument has not been scalarized
1951 if (_gvn.type(arg)->is_zero_type()) {
1952 arg = InlineTypeNode::make_null(_gvn, t->inline_klass());
1953 } else {
1954 // During parsing, a method is called with an abstract (or j.l.Object) receiver, the
1955 // receiver is a non-scalarized oop. Later on, IGVN reveals that the receiver must be a
1956 // value object. The method is devirtualized, and replaced with a direct call with a
1957 // scalarized receiver instead.
1958 assert(arg_idx == 0 && !call->method()->is_static(), "must be the receiver");
1959 assert(C->inlining_incrementally() || C->strength_reduction(), "must be during devirtualization of calls");
1960 assert(!is_Parse(), "must be during devirtualization of calls");
1961 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
1962 }
1963 }
1964 InlineTypeNode* vt = arg->as_InlineType();
1965 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1966 // If an inline type argument is passed as fields, attach the Method* to the call site
1967 // to be able to access the extended signature later via attached_method_before_pc().
1968 // For example, see CompiledMethod::preserve_callee_argument_oops().
1969 call->set_override_symbolic_info(true);
1970 // Register an calling convention dependency on the callee method to make sure that this method is deoptimized and
1971 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1972 C->dependencies()->assert_mismatch_calling_convention(call->method());
1973 arg_num++;
1974 continue;
1975 } else if (arg->is_InlineType()) {
1976 // Pass inline type argument via oop to callee
1977 arg = arg->as_InlineType()->buffer(this, true);
1978 }
1979 if (t != Type::HALF) {
1980 arg_num++;
1981 }
1982 call->init_req(idx++, arg);
1983 }
1984 }
1985
1986 //---------------------------set_edges_for_java_call---------------------------
1987 // Connect a newly created call into the current JVMS.
1988 // A return value node (if any) is returned from set_edges_for_java_call.
1989 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1990
1991 // Add the predefined inputs:
1992 call->init_req( TypeFunc::Control, control() );
1993 call->init_req( TypeFunc::I_O , i_o() );
1994 call->init_req( TypeFunc::Memory , reset_memory() );
1995 call->init_req( TypeFunc::FramePtr, frameptr() );
1996 call->init_req( TypeFunc::ReturnAdr, top() );
1997
1998 add_safepoint_edges(call, must_throw);
1999
2000 Node* xcall = _gvn.transform(call);
2001
2002 if (xcall == top()) {
2003 set_control(top());
2004 return;
2005 }
2006 assert(xcall == call, "call identity is stable");
2007
2008 // Re-use the current map to produce the result.
2009
2010 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
2011 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
2012 set_all_memory_call(xcall, separate_io_proj);
2013
2014 //return xcall; // no need, caller already has it
2015 }
2016
2017 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2018 if (stopped()) return top(); // maybe the call folded up?
2019
2020 // Note: Since any out-of-line call can produce an exception,
2021 // we always insert an I_O projection from the call into the result.
2022
2023 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2024
2025 if (separate_io_proj) {
2026 // The caller requested separate projections be used by the fall
2027 // through and exceptional paths, so replace the projections for
2028 // the fall through path.
2029 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2030 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2031 }
2032
2033 // Capture the return value, if any.
2034 Node* ret;
2035 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2036 ret = top();
2037 } else if (call->tf()->returns_inline_type_as_fields()) {
2038 // Return of multiple values (inline type fields): we create a
2039 // InlineType node, each field is a projection from the call.
2040 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2041 uint base_input = TypeFunc::Parms;
2042 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2043 } else {
2044 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2045 ciType* t = call->method()->return_type();
2046 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2047 // The return type is unloaded but the callee might later be C2 compiled and then return
2048 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2049 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2050 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2051 IdealKit ideal(this);
2052 IdealVariable res(ideal);
2053 ideal.declarations_done();
2054 // Change return type of call to scalarized return
2055 const TypeFunc* tf = call->_tf;
2056 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2057 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2058 call->_tf = new_tf;
2059 _gvn.set_type(call, call->Value(&_gvn));
2060 _gvn.set_type(ret, ret->Value(&_gvn));
2061 // Don't add store to buffer call if we are strength reducing
2062 if (!C->strength_reduction()) {
2063 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2064 // Return value is null
2065 ideal.set(res, makecon(TypePtr::NULL_PTR));
2066 } ideal.else_(); {
2067 // Return value is non-null
2068 sync_kit(ideal);
2069
2070 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2071 OptoRuntime::store_inline_type_fields_Type(),
2072 StubRoutines::store_inline_type_fields_to_buf(),
2073 nullptr, TypePtr::BOTTOM, ret);
2074
2075 // We don't know how many values are returned. This assumes the
2076 // worst case, that all available registers are used.
2077 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2078 if (domain->field_at(i) == Type::HALF) {
2079 store_to_buf_call->init_req(i, top());
2080 continue;
2081 }
2082 Node* proj =_gvn.transform(new ProjNode(call, i));
2083 store_to_buf_call->init_req(i, proj);
2084 }
2085 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2086
2087 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2088 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2089 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2090
2091 ideal.set(res, buf);
2092 ideal.sync_kit(this);
2093 } ideal.end_if();
2094 } else {
2095 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2096 Node* proj =_gvn.transform(new ProjNode(call, i));
2097 }
2098 ideal.set(res, ret);
2099 }
2100 sync_kit(ideal);
2101 ret = _gvn.transform(ideal.value(res));
2102 }
2103 if (t->is_klass()) {
2104 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
2105 if (type->is_inlinetypeptr()) {
2106 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass());
2107 }
2108 }
2109 }
2110
2111 return ret;
2112 }
2113
2114 //--------------------set_predefined_input_for_runtime_call--------------------
2115 // Reading and setting the memory state is way conservative here.
2116 // The real problem is that I am not doing real Type analysis on memory,
2117 // so I cannot distinguish card mark stores from other stores. Across a GC
2118 // point the Store Barrier and the card mark memory has to agree. I cannot
2119 // have a card mark store and its barrier split across the GC point from
2120 // either above or below. Here I get that to happen by reading ALL of memory.
2121 // A better answer would be to separate out card marks from other memory.
2122 // For now, return the input memory state, so that it can be reused
2123 // after the call, if this call has restricted memory effects.
2124 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2125 // Set fixed predefined input arguments
2126 call->init_req(TypeFunc::Control, control());
2127 call->init_req(TypeFunc::I_O, top()); // does no i/o
2128 call->init_req(TypeFunc::ReturnAdr, top());
2129 if (call->is_CallLeafPure()) {
2130 call->init_req(TypeFunc::Memory, top());
2192 if (use->is_MergeMem()) {
2193 wl.push(use);
2194 }
2195 }
2196 }
2197
2198 // Replace the call with the current state of the kit.
2199 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2200 JVMState* ejvms = nullptr;
2201 if (has_exceptions()) {
2202 ejvms = transfer_exceptions_into_jvms();
2203 }
2204
2205 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2206 ReplacedNodes replaced_nodes_exception;
2207 Node* ex_ctl = top();
2208
2209 SafePointNode* final_state = stop();
2210
2211 // Find all the needed outputs of this call
2212 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2213
2214 Unique_Node_List wl;
2215 Node* init_mem = call->in(TypeFunc::Memory);
2216 Node* final_mem = final_state->in(TypeFunc::Memory);
2217 Node* final_ctl = final_state->in(TypeFunc::Control);
2218 Node* final_io = final_state->in(TypeFunc::I_O);
2219
2220 // Replace all the old call edges with the edges from the inlining result
2221 if (callprojs->fallthrough_catchproj != nullptr) {
2222 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2223 }
2224 if (callprojs->fallthrough_memproj != nullptr) {
2225 if (final_mem->is_MergeMem()) {
2226 // Parser's exits MergeMem was not transformed but may be optimized
2227 final_mem = _gvn.transform(final_mem);
2228 }
2229 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2230 add_mergemem_users_to_worklist(wl, final_mem);
2231 }
2232 if (callprojs->fallthrough_ioproj != nullptr) {
2233 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2234 }
2235
2236 // Replace the result with the new result if it exists and is used
2237 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2238 // If the inlined code is dead, the result projections for an inline type returned as
2239 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2240 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()) ||
2241 (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()),
2242 "unexpected number of results");
2243 // If we are doing strength reduction and the return type is not loaded we
2244 // need to rewire all projections since store_inline_type_fields_to_buf is already present
2245 if (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()) {
2246 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2247 for (uint i = TypeFunc::Parms; i < domain->cnt(); i++) {
2248 C->gvn_replace_by(callprojs->resproj[0], final_state->in(i));
2249 }
2250 } else {
2251 C->gvn_replace_by(callprojs->resproj[0], result);
2252 }
2253 }
2254
2255 if (ejvms == nullptr) {
2256 // No exception edges to simply kill off those paths
2257 if (callprojs->catchall_catchproj != nullptr) {
2258 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2259 }
2260 if (callprojs->catchall_memproj != nullptr) {
2261 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2262 }
2263 if (callprojs->catchall_ioproj != nullptr) {
2264 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2265 }
2266 // Replace the old exception object with top
2267 if (callprojs->exobj != nullptr) {
2268 C->gvn_replace_by(callprojs->exobj, C->top());
2269 }
2270 } else {
2271 GraphKit ekit(ejvms);
2272
2273 // Load my combined exception state into the kit, with all phis transformed:
2274 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2275 replaced_nodes_exception = ex_map->replaced_nodes();
2276
2277 Node* ex_oop = ekit.use_exception_state(ex_map);
2278
2279 if (callprojs->catchall_catchproj != nullptr) {
2280 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2281 ex_ctl = ekit.control();
2282 }
2283 if (callprojs->catchall_memproj != nullptr) {
2284 Node* ex_mem = ekit.reset_memory();
2285 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2286 add_mergemem_users_to_worklist(wl, ex_mem);
2287 }
2288 if (callprojs->catchall_ioproj != nullptr) {
2289 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2290 }
2291
2292 // Replace the old exception object with the newly created one
2293 if (callprojs->exobj != nullptr) {
2294 C->gvn_replace_by(callprojs->exobj, ex_oop);
2295 }
2296 }
2297
2298 // Disconnect the call from the graph
2299 call->disconnect_inputs(C);
2300 C->gvn_replace_by(call, C->top());
2301
2302 // Clean up any MergeMems that feed other MergeMems since the
2303 // optimizer doesn't like that.
2304 while (wl.size() > 0) {
2305 _gvn.transform(wl.pop());
2306 }
2307
2308 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2309 replaced_nodes.apply(C, final_ctl);
2310 }
2311 if (!ex_ctl->is_top() && do_replaced_nodes) {
2312 replaced_nodes_exception.apply(C, ex_ctl);
2313 }
2314 }
2315
2316
2317 //------------------------------increment_counter------------------------------
2318 // for statistics: increment a VM counter by 1
2319
2320 void GraphKit::increment_counter(address counter_addr) {
2321 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2322 increment_counter(adr1);
2323 }
2324
2325 void GraphKit::increment_counter(Node* counter_addr) {
2326 Node* ctrl = control();
2327 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2328 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2498 *
2499 * @param n node that the type applies to
2500 * @param exact_kls type from profiling
2501 * @param maybe_null did profiling see null?
2502 *
2503 * @return node with improved type
2504 */
2505 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2506 const Type* current_type = _gvn.type(n);
2507 assert(UseTypeSpeculation, "type speculation must be on");
2508
2509 const TypePtr* speculative = current_type->speculative();
2510
2511 // Should the klass from the profile be recorded in the speculative type?
2512 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2513 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2514 const TypeOopPtr* xtype = tklass->as_instance_type();
2515 assert(xtype->klass_is_exact(), "Should be exact");
2516 // Any reason to believe n is not null (from this profiling or a previous one)?
2517 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2518 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2519 // record the new speculative type's depth
2520 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2521 speculative = speculative->with_inline_depth(jvms()->depth());
2522 } else if (current_type->would_improve_ptr(ptr_kind)) {
2523 // Profiling report that null was never seen so we can change the
2524 // speculative type to non null ptr.
2525 if (ptr_kind == ProfileAlwaysNull) {
2526 speculative = TypePtr::NULL_PTR;
2527 } else {
2528 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2529 const TypePtr* ptr = TypePtr::NOTNULL;
2530 if (speculative != nullptr) {
2531 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2532 } else {
2533 speculative = ptr;
2534 }
2535 }
2536 }
2537
2538 if (speculative != current_type->speculative()) {
2539 // Build a type with a speculative type (what we think we know
2540 // about the type but will need a guard when we use it)
2541 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2542 // We're changing the type, we need a new CheckCast node to carry
2543 // the new type. The new type depends on the control: what
2544 // profiling tells us is only valid from here as far as we can
2545 // tell.
2546 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2547 cast = _gvn.transform(cast);
2548 replace_in_map(n, cast);
2549 n = cast;
2550 }
2551
2552 return n;
2553 }
2554
2555 /**
2556 * Record profiling data from receiver profiling at an invoke with the
2557 * type system so that it can propagate it (speculation)
2558 *
2559 * @param n receiver node
2560 *
2561 * @return node with improved type
2562 */
2563 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2564 if (!UseTypeSpeculation) {
2565 return n;
2566 }
2567 ciKlass* exact_kls = profile_has_unique_klass();
2568 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2569 if ((java_bc() == Bytecodes::_checkcast ||
2570 java_bc() == Bytecodes::_instanceof ||
2571 java_bc() == Bytecodes::_aastore) &&
2572 method()->method_data()->is_mature()) {
2573 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2574 if (data != nullptr) {
2575 if (java_bc() == Bytecodes::_aastore) {
2576 ciKlass* array_type = nullptr;
2577 ciKlass* element_type = nullptr;
2578 ProfilePtrKind element_ptr = ProfileMaybeNull;
2579 bool flat_array = true;
2580 bool null_free_array = true;
2581 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2582 exact_kls = element_type;
2583 ptr_kind = element_ptr;
2584 } else {
2585 if (!data->as_BitData()->null_seen()) {
2586 ptr_kind = ProfileNeverNull;
2587 } else {
2588 if (TypeProfileCasts) {
2589 assert(data->is_ReceiverTypeData(), "bad profile data type");
2590 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2591 uint i = 0;
2592 for (; i < call->row_limit(); i++) {
2593 ciKlass* receiver = call->receiver(i);
2594 if (receiver != nullptr) {
2595 break;
2596 }
2597 }
2598 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2599 }
2600 }
2601 }
2602 }
2603 }
2604 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2605 }
2606
2607 /**
2608 * Record profiling data from argument profiling at an invoke with the
2609 * type system so that it can propagate it (speculation)
2610 *
2611 * @param dest_method target method for the call
2612 * @param bc what invoke bytecode is this?
2613 */
2614 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2615 if (!UseTypeSpeculation) {
2616 return;
2617 }
2618 const TypeFunc* tf = TypeFunc::make(dest_method);
2619 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2620 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2621 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2622 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2623 if (is_reference_type(targ->basic_type())) {
2624 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2625 ciKlass* better_type = nullptr;
2626 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2627 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2628 }
2629 i++;
2630 }
2631 }
2632 }
2633
2634 /**
2635 * Record profiling data from parameter profiling at an invoke with
2636 * the type system so that it can propagate it (speculation)
2637 */
2638 void GraphKit::record_profiled_parameters_for_speculation() {
2639 if (!UseTypeSpeculation) {
2640 return;
2641 }
2642 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2762 // The first null ends the list.
2763 Node* parm0, Node* parm1,
2764 Node* parm2, Node* parm3,
2765 Node* parm4, Node* parm5,
2766 Node* parm6, Node* parm7) {
2767 assert(call_addr != nullptr, "must not call null targets");
2768
2769 // Slow-path call
2770 bool is_leaf = !(flags & RC_NO_LEAF);
2771 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2772 if (call_name == nullptr) {
2773 assert(!is_leaf, "must supply name for leaf");
2774 call_name = OptoRuntime::stub_name(call_addr);
2775 }
2776 CallNode* call;
2777 if (!is_leaf) {
2778 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2779 } else if (flags & RC_NO_FP) {
2780 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2781 } else if (flags & RC_VECTOR){
2782 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2783 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2784 } else if (flags & RC_PURE) {
2785 assert(adr_type == nullptr, "pure call does not touch memory");
2786 call = new CallLeafPureNode(call_type, call_addr, call_name);
2787 } else {
2788 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2789 }
2790
2791 // The following is similar to set_edges_for_java_call,
2792 // except that the memory effects of the call are restricted to AliasIdxRaw.
2793
2794 // Slow path call has no side-effects, uses few values
2795 bool wide_in = !(flags & RC_NARROW_MEM);
2796 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2797
2798 Node* prev_mem = nullptr;
2799 if (wide_in) {
2800 prev_mem = set_predefined_input_for_runtime_call(call);
2801 } else {
2802 assert(!wide_out, "narrow in => narrow out");
2803 Node* narrow_mem = memory(adr_type);
2804 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2805 }
2806
2807 // Hook each parm in order. Stop looking at the first null.
2808 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2809 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2810 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2811 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2812 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2813 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2814 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2815 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2816 /* close each nested if ===> */ } } } } } } } }
2817 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2818
2819 if (!is_leaf) {
2820 // Non-leaves can block and take safepoints:
2821 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2822 }
2823 // Non-leaves can throw exceptions:
2824 if (has_io) {
2825 call->set_req(TypeFunc::I_O, i_o());
2826 }
2827
2828 if (flags & RC_UNCOMMON) {
2829 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2830 // (An "if" probability corresponds roughly to an unconditional count.
2831 // Sort of.)
2832 call->set_cnt(PROB_UNLIKELY_MAG(4));
2833 }
2834
2835 Node* c = _gvn.transform(call);
2836 assert(c == call, "cannot disappear");
2837
2845
2846 if (has_io) {
2847 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2848 }
2849 return call;
2850
2851 }
2852
2853 // i2b
2854 Node* GraphKit::sign_extend_byte(Node* in) {
2855 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2856 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2857 }
2858
2859 // i2s
2860 Node* GraphKit::sign_extend_short(Node* in) {
2861 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2862 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2863 }
2864
2865
2866 //------------------------------merge_memory-----------------------------------
2867 // Merge memory from one path into the current memory state.
2868 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2869 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2870 Node* old_slice = mms.force_memory();
2871 Node* new_slice = mms.memory2();
2872 if (old_slice != new_slice) {
2873 PhiNode* phi;
2874 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2875 if (mms.is_empty()) {
2876 // clone base memory Phi's inputs for this memory slice
2877 assert(old_slice == mms.base_memory(), "sanity");
2878 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2879 _gvn.set_type(phi, Type::MEMORY);
2880 for (uint i = 1; i < phi->req(); i++) {
2881 phi->init_req(i, old_slice->in(i));
2882 }
2883 } else {
2884 phi = old_slice->as_Phi(); // Phi was generated already
2885 }
2942 gvn.transform(iff);
2943 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2944 return iff;
2945 }
2946
2947 //-------------------------------gen_subtype_check-----------------------------
2948 // Generate a subtyping check. Takes as input the subtype and supertype.
2949 // Returns 2 values: sets the default control() to the true path and returns
2950 // the false path. Only reads invariant memory; sets no (visible) memory.
2951 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2952 // but that's not exposed to the optimizer. This call also doesn't take in an
2953 // Object; if you wish to check an Object you need to load the Object's class
2954 // prior to coming here.
2955 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2956 ciMethod* method, int bci) {
2957 Compile* C = gvn.C;
2958 if ((*ctrl)->is_top()) {
2959 return C->top();
2960 }
2961
2962 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
2963 // For a direct pointer comparison, we need the refined array klass pointer
2964 Node* vm_superklass = superklass;
2965 if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
2966 assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
2967 vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
2968 }
2969
2970 // Fast check for identical types, perhaps identical constants.
2971 // The types can even be identical non-constants, in cases
2972 // involving Array.newInstance, Object.clone, etc.
2973 if (subklass == superklass)
2974 return C->top(); // false path is dead; no test needed.
2975
2976 if (gvn.type(superklass)->singleton()) {
2977 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2978 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2979
2980 // In the common case of an exact superklass, try to fold up the
2981 // test before generating code. You may ask, why not just generate
2982 // the code and then let it fold up? The answer is that the generated
2983 // code will necessarily include null checks, which do not always
2984 // completely fold away. If they are also needless, then they turn
2985 // into a performance loss. Example:
2986 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2987 // Here, the type of 'fa' is often exact, so the store check
2988 // of fa[1]=x will fold up, without testing the nullness of x.
2989 //
2990 // At macro expansion, we would have already folded the SubTypeCheckNode
2991 // being expanded here because we always perform the static sub type
2992 // check in SubTypeCheckNode::sub() regardless of whether
2993 // StressReflectiveCode is set or not. We can therefore skip this
2994 // static check when StressReflectiveCode is on.
2995 switch (C->static_subtype_check(superk, subk)) {
2996 case Compile::SSC_always_false:
2997 {
2998 Node* always_fail = *ctrl;
2999 *ctrl = gvn.C->top();
3000 return always_fail;
3001 }
3002 case Compile::SSC_always_true:
3003 return C->top();
3004 case Compile::SSC_easy_test:
3005 {
3006 // Just do a direct pointer compare and be done.
3007 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
3008 *ctrl = gvn.transform(new IfTrueNode(iff));
3009 return gvn.transform(new IfFalseNode(iff));
3010 }
3011 case Compile::SSC_full_test:
3012 break;
3013 default:
3014 ShouldNotReachHere();
3015 }
3016 }
3017
3018 // %%% Possible further optimization: Even if the superklass is not exact,
3019 // if the subklass is the unique subtype of the superklass, the check
3020 // will always succeed. We could leave a dependency behind to ensure this.
3021
3022 // First load the super-klass's check-offset
3023 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
3024 Node* m = C->immutable_memory();
3025 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
3026 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
3027 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
3065 gvn.record_for_igvn(r_ok_subtype);
3066
3067 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3068 // SubTypeCheck node
3069 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3070 ciCallProfile profile = method->call_profile_at_bci(bci);
3071 float total_prob = 0;
3072 for (int i = 0; profile.has_receiver(i); ++i) {
3073 float prob = profile.receiver_prob(i);
3074 total_prob += prob;
3075 }
3076 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3077 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3078 for (int i = 0; profile.has_receiver(i); ++i) {
3079 ciKlass* klass = profile.receiver(i);
3080 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3081 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3082 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3083 continue;
3084 }
3085 if (klass_t->isa_aryklassptr()) {
3086 // For a direct pointer comparison, we need the refined array klass pointer
3087 klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3088 }
3089 float prob = profile.receiver_prob(i);
3090 ConNode* klass_node = gvn.makecon(klass_t);
3091 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3092 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3093
3094 if (result == Compile::SSC_always_true) {
3095 r_ok_subtype->add_req(iftrue);
3096 } else {
3097 assert(result == Compile::SSC_always_false, "");
3098 r_not_subtype->add_req(iftrue);
3099 }
3100 *ctrl = gvn.transform(new IfFalseNode(iff));
3101 }
3102 }
3103 }
3104
3105 // See if we get an immediate positive hit. Happens roughly 83% of the
3106 // time. Test to see if the value loaded just previously from the subklass
3107 // is exactly the superklass.
3108 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3122 igvn->remove_globally_dead_node(r_not_subtype);
3123 }
3124 return not_subtype_ctrl;
3125 }
3126
3127 r_ok_subtype->init_req(1, iftrue1);
3128
3129 // Check for immediate negative hit. Happens roughly 11% of the time (which
3130 // is roughly 63% of the remaining cases). Test to see if the loaded
3131 // check-offset points into the subklass display list or the 1-element
3132 // cache. If it points to the display (and NOT the cache) and the display
3133 // missed then it's not a subtype.
3134 Node *cacheoff = gvn.intcon(cacheoff_con);
3135 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3136 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3137 *ctrl = gvn.transform(new IfFalseNode(iff2));
3138
3139 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3140 // No performance impact (too rare) but allows sharing of secondary arrays
3141 // which has some footprint reduction.
3142 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3143 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3144 *ctrl = gvn.transform(new IfFalseNode(iff3));
3145
3146 // -- Roads not taken here: --
3147 // We could also have chosen to perform the self-check at the beginning
3148 // of this code sequence, as the assembler does. This would not pay off
3149 // the same way, since the optimizer, unlike the assembler, can perform
3150 // static type analysis to fold away many successful self-checks.
3151 // Non-foldable self checks work better here in second position, because
3152 // the initial primary superclass check subsumes a self-check for most
3153 // types. An exception would be a secondary type like array-of-interface,
3154 // which does not appear in its own primary supertype display.
3155 // Finally, we could have chosen to move the self-check into the
3156 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3157 // dependent manner. But it is worthwhile to have the check here,
3158 // where it can be perhaps be optimized. The cost in code space is
3159 // small (register compare, branch).
3160
3161 // Now do a linear scan of the secondary super-klass array. Again, no real
3162 // performance impact (too rare) but it's gotta be done.
3163 // Since the code is rarely used, there is no penalty for moving it
3164 // out of line, and it can only improve I-cache density.
3165 // The decision to inline or out-of-line this final check is platform
3166 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3167 Node* psc = gvn.transform(
3168 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3169
3170 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3171 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3172 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3173
3174 // Return false path; set default control to true path.
3175 *ctrl = gvn.transform(r_ok_subtype);
3176 return gvn.transform(r_not_subtype);
3177 }
3178
3179 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3180 const Type* sub_t = _gvn.type(obj_or_subklass);
3181 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3182 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3183 obj_or_subklass = makecon(sub_t);
3184 }
3185 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3186 if (expand_subtype_check) {
3187 MergeMemNode* mem = merged_memory();
3188 Node* ctrl = control();
3189 Node* subklass = obj_or_subklass;
3190 if (!sub_t->isa_klassptr()) {
3191 subklass = load_object_klass(obj_or_subklass);
3192 }
3193
3194 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3195 set_control(ctrl);
3196 return n;
3197 }
3198
3199 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3200 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3201 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3202 set_control(_gvn.transform(new IfTrueNode(iff)));
3203 return _gvn.transform(new IfFalseNode(iff));
3204 }
3205
3206 // Profile-driven exact type check:
3207 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3208 float prob, Node* *casted_receiver) {
3209 assert(!klass->is_interface(), "no exact type check on interfaces");
3210 Node* fail = top();
3211 const Type* rec_t = _gvn.type(receiver);
3212 if (rec_t->is_inlinetypeptr()) {
3213 if (klass->equals(rec_t->inline_klass())) {
3214 (*casted_receiver) = receiver; // Always passes
3215 } else {
3216 (*casted_receiver) = top(); // Always fails
3217 fail = control();
3218 set_control(top());
3219 }
3220 return fail;
3221 }
3222 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3223 if (tklass->isa_aryklassptr()) {
3224 // For a direct pointer comparison, we need the refined array klass pointer
3225 tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3226 }
3227 Node* recv_klass = load_object_klass(receiver);
3228 fail = type_check(recv_klass, tklass, prob);
3229
3230 if (!stopped()) {
3231 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3232 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3233 assert(recv_xtype->klass_is_exact(), "");
3234
3235 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3236 // Subsume downstream occurrences of receiver with a cast to
3237 // recv_xtype, since now we know what the type will be.
3238 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3239 Node* res = _gvn.transform(cast);
3240 if (recv_xtype->is_inlinetypeptr()) {
3241 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3242 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3243 }
3244 (*casted_receiver) = res;
3245 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3246 // (User must make the replace_in_map call.)
3247 }
3248 }
3249
3250 return fail;
3251 }
3252
3253 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3254 float prob) {
3255 Node* want_klass = makecon(tklass);
3256 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3257 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3258 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3259 set_control(_gvn.transform(new IfTrueNode (iff)));
3260 Node* fail = _gvn.transform(new IfFalseNode(iff));
3261 return fail;
3262 }
3263
3264 //------------------------------subtype_check_receiver-------------------------
3265 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3266 Node** casted_receiver) {
3267 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3268 Node* want_klass = makecon(tklass);
3269
3270 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3271
3272 // Ignore interface type information until interface types are properly tracked.
3273 if (!stopped() && !klass->is_interface()) {
3274 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3275 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3276 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3277 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3278 if (recv_type->is_inlinetypeptr()) {
3279 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3280 }
3281 (*casted_receiver) = cast;
3282 }
3283 }
3284
3285 return slow_ctl;
3286 }
3287
3288 //------------------------------seems_never_null-------------------------------
3289 // Use null_seen information if it is available from the profile.
3290 // If we see an unexpected null at a type check we record it and force a
3291 // recompile; the offending check will be recompiled to handle nulls.
3292 // If we see several offending BCIs, then all checks in the
3293 // method will be recompiled.
3294 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3295 speculating = !_gvn.type(obj)->speculative_maybe_null();
3296 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3297 if (UncommonNullCast // Cutout for this technique
3298 && obj != null() // And not the -Xcomp stupid case?
3299 && !too_many_traps(reason)
3300 ) {
3301 if (speculating) {
3370
3371 //------------------------maybe_cast_profiled_receiver-------------------------
3372 // If the profile has seen exactly one type, narrow to exactly that type.
3373 // Subsequent type checks will always fold up.
3374 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3375 const TypeKlassPtr* require_klass,
3376 ciKlass* spec_klass,
3377 bool safe_for_replace) {
3378 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3379
3380 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3381
3382 // Make sure we haven't already deoptimized from this tactic.
3383 if (too_many_traps_or_recompiles(reason))
3384 return nullptr;
3385
3386 // (No, this isn't a call, but it's enough like a virtual call
3387 // to use the same ciMethod accessor to get the profile info...)
3388 // If we have a speculative type use it instead of profiling (which
3389 // may not help us)
3390 ciKlass* exact_kls = spec_klass;
3391 if (exact_kls == nullptr) {
3392 if (java_bc() == Bytecodes::_aastore) {
3393 ciKlass* array_type = nullptr;
3394 ciKlass* element_type = nullptr;
3395 ProfilePtrKind element_ptr = ProfileMaybeNull;
3396 bool flat_array = true;
3397 bool null_free_array = true;
3398 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3399 exact_kls = element_type;
3400 } else {
3401 exact_kls = profile_has_unique_klass();
3402 }
3403 }
3404 if (exact_kls != nullptr) {// no cast failures here
3405 if (require_klass == nullptr ||
3406 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3407 // If we narrow the type to match what the type profile sees or
3408 // the speculative type, we can then remove the rest of the
3409 // cast.
3410 // This is a win, even if the exact_kls is very specific,
3411 // because downstream operations, such as method calls,
3412 // will often benefit from the sharper type.
3413 Node* exact_obj = not_null_obj; // will get updated in place...
3414 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3415 &exact_obj);
3416 { PreserveJVMState pjvms(this);
3417 set_control(slow_ctl);
3418 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3419 }
3420 if (safe_for_replace) {
3421 replace_in_map(not_null_obj, exact_obj);
3422 }
3423 return exact_obj;
3513 // If not_null_obj is dead, only null-path is taken
3514 if (stopped()) { // Doing instance-of on a null?
3515 set_control(null_ctl);
3516 return intcon(0);
3517 }
3518 region->init_req(_null_path, null_ctl);
3519 phi ->init_req(_null_path, intcon(0)); // Set null path value
3520 if (null_ctl == top()) {
3521 // Do this eagerly, so that pattern matches like is_diamond_phi
3522 // will work even during parsing.
3523 assert(_null_path == PATH_LIMIT-1, "delete last");
3524 region->del_req(_null_path);
3525 phi ->del_req(_null_path);
3526 }
3527
3528 // Do we know the type check always succeed?
3529 bool known_statically = false;
3530 if (_gvn.type(superklass)->singleton()) {
3531 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3532 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3533 if (subk != nullptr && subk->is_loaded()) {
3534 int static_res = C->static_subtype_check(superk, subk);
3535 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3536 }
3537 }
3538
3539 if (!known_statically) {
3540 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3541 // We may not have profiling here or it may not help us. If we
3542 // have a speculative type use it to perform an exact cast.
3543 ciKlass* spec_obj_type = obj_type->speculative_type();
3544 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3545 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3546 if (stopped()) { // Profile disagrees with this path.
3547 set_control(null_ctl); // Null is the only remaining possibility.
3548 return intcon(0);
3549 }
3550 if (cast_obj != nullptr) {
3551 not_null_obj = cast_obj;
3552 }
3553 }
3569 record_for_igvn(region);
3570
3571 // If we know the type check always succeeds then we don't use the
3572 // profiling data at this bytecode. Don't lose it, feed it to the
3573 // type system as a speculative type.
3574 if (safe_for_replace) {
3575 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3576 replace_in_map(obj, casted_obj);
3577 }
3578
3579 return _gvn.transform(phi);
3580 }
3581
3582 //-------------------------------gen_checkcast---------------------------------
3583 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3584 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3585 // uncommon-trap paths work. Adjust stack after this call.
3586 // If failure_control is supplied and not null, it is filled in with
3587 // the control edge for the cast failure. Otherwise, an appropriate
3588 // uncommon trap or exception is thrown.
3589 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free, bool maybe_larval) {
3590 kill_dead_locals(); // Benefit all the uncommon traps
3591 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3592 const Type* obj_type = _gvn.type(obj);
3593 obj = cast_to_non_larval(obj);
3594
3595 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3596 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3597 bool safe_for_replace = (failure_control == nullptr);
3598 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3599
3600 // Fast cutout: Check the case that the cast is vacuously true.
3601 // This detects the common cases where the test will short-circuit
3602 // away completely. We do this before we perform the null check,
3603 // because if the test is going to turn into zero code, we don't
3604 // want a residual null check left around. (Causes a slowdown,
3605 // for example, in some objArray manipulations, such as a[i]=a[j].)
3606 if (improved_klass_ptr_type->singleton()) {
3607 const TypeKlassPtr* kptr = nullptr;
3608 if (obj_type->isa_oop_ptr()) {
3609 kptr = obj_type->is_oopptr()->as_klass_type();
3610 } else if (obj->is_InlineType()) {
3611 ciInlineKlass* vk = obj_type->inline_klass();
3612 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3613 }
3614
3615 if (kptr != nullptr) {
3616 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3617 case Compile::SSC_always_true:
3618 // If we know the type check always succeed then we don't use
3619 // the profiling data at this bytecode. Don't lose it, feed it
3620 // to the type system as a speculative type.
3621 obj = record_profiled_receiver_for_speculation(obj);
3622 if (null_free) {
3623 assert(safe_for_replace, "must be");
3624 obj = null_check(obj);
3625 }
3626 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3627 return obj;
3628 case Compile::SSC_always_false:
3629 if (null_free) {
3630 assert(safe_for_replace, "must be");
3631 obj = null_check(obj);
3632 }
3633 // It needs a null check because a null will *pass* the cast check.
3634 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3635 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3636 Deoptimization::DeoptReason reason = is_aastore ?
3637 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3638 builtin_throw(reason);
3639 return top();
3640 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3641 return null_assert(obj);
3642 }
3643 break; // Fall through to full check
3644 default:
3645 break;
3646 }
3647 }
3648 }
3649
3650 ciProfileData* data = nullptr;
3651 if (failure_control == nullptr) { // use MDO in regular case only
3652 assert(java_bc() == Bytecodes::_aastore ||
3653 java_bc() == Bytecodes::_checkcast,
3654 "interpreter profiles type checks only for these BCs");
3655 if (method()->method_data()->is_mature()) {
3656 data = method()->method_data()->bci_to_data(bci());
3657 }
3658 }
3659
3660 // Make the merge point
3661 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3662 RegionNode* region = new RegionNode(PATH_LIMIT);
3663 Node* phi = new PhiNode(region, toop);
3664 _gvn.set_type(region, Type::CONTROL);
3665 _gvn.set_type(phi, toop);
3666
3667 C->set_has_split_ifs(true); // Has chance for split-if optimization
3668
3669 // Use null-cast information if it is available
3670 bool speculative_not_null = false;
3671 bool never_see_null = ((failure_control == nullptr) // regular case only
3672 && seems_never_null(obj, data, speculative_not_null));
3673
3674 if (obj->is_InlineType()) {
3675 // Re-execute if buffering during triggers deoptimization
3676 PreserveReexecuteState preexecs(this);
3677 jvms()->set_should_reexecute(true);
3678 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3679 }
3680
3681 // Null check; get casted pointer; set region slot 3
3682 Node* null_ctl = top();
3683 Node* not_null_obj = nullptr;
3684 if (null_free) {
3685 assert(safe_for_replace, "must be");
3686 not_null_obj = null_check(obj);
3687 } else {
3688 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3689 }
3690
3691 // If not_null_obj is dead, only null-path is taken
3692 if (stopped()) { // Doing instance-of on a null?
3693 set_control(null_ctl);
3694 if (toop->is_inlinetypeptr()) {
3695 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3696 }
3697 return null();
3698 }
3699 region->init_req(_null_path, null_ctl);
3700 phi ->init_req(_null_path, null()); // Set null path value
3701 if (null_ctl == top()) {
3702 // Do this eagerly, so that pattern matches like is_diamond_phi
3703 // will work even during parsing.
3704 assert(_null_path == PATH_LIMIT-1, "delete last");
3705 region->del_req(_null_path);
3706 phi ->del_req(_null_path);
3707 }
3708
3709 Node* cast_obj = nullptr;
3710 if (improved_klass_ptr_type->klass_is_exact()) {
3711 // The following optimization tries to statically cast the speculative type of the object
3712 // (for example obtained during profiling) to the type of the superklass and then do a
3713 // dynamic check that the type of the object is what we expect. To work correctly
3714 // for checkcast and aastore the type of superklass should be exact.
3715 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3716 // We may not have profiling here or it may not help us. If we have
3717 // a speculative type use it to perform an exact cast.
3718 ciKlass* spec_obj_type = obj_type->speculative_type();
3719 if (spec_obj_type != nullptr || data != nullptr) {
3720 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3721 if (cast_obj != nullptr) {
3722 if (failure_control != nullptr) // failure is now impossible
3723 (*failure_control) = top();
3724 // adjust the type of the phi to the exact klass:
3725 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3726 }
3727 }
3728 }
3729
3730 if (cast_obj == nullptr) {
3731 // Generate the subtype check
3732 Node* improved_superklass = superklass;
3733 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3734 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3735 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3736 // Additionally, the benefit would only be minor in non-constant cases.
3737 improved_superklass = makecon(improved_klass_ptr_type);
3738 }
3739 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3740 // Plug in success path into the merge
3741 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3742 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3743 if (failure_control == nullptr) {
3744 if (not_subtype_ctrl != top()) { // If failure is possible
3745 PreserveJVMState pjvms(this);
3746 set_control(not_subtype_ctrl);
3747 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3748 Deoptimization::DeoptReason reason = is_aastore ?
3749 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3750 builtin_throw(reason);
3751 }
3752 } else {
3753 (*failure_control) = not_subtype_ctrl;
3754 }
3755 }
3756
3757 region->init_req(_obj_path, control());
3758 phi ->init_req(_obj_path, cast_obj);
3759
3760 // A merge of null or Casted-NotNull obj
3761 Node* res = _gvn.transform(phi);
3762
3763 // Note I do NOT always 'replace_in_map(obj,result)' here.
3764 // if( tk->klass()->can_be_primary_super() )
3765 // This means that if I successfully store an Object into an array-of-String
3766 // I 'forget' that the Object is really now known to be a String. I have to
3767 // do this because we don't have true union types for interfaces - if I store
3768 // a Baz into an array-of-Interface and then tell the optimizer it's an
3769 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3770 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3771 // replace_in_map( obj, res );
3772
3773 // Return final merged results
3774 set_control( _gvn.transform(region) );
3775 record_for_igvn(region);
3776
3777 bool not_inline = !toop->can_be_inline_type();
3778 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3779 if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3780 // Check if obj has been loaded from an array
3781 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3782 Node* array = nullptr;
3783 if (obj->isa_Load()) {
3784 Node* address = obj->in(MemNode::Address);
3785 if (address->isa_AddP()) {
3786 array = address->as_AddP()->in(AddPNode::Base);
3787 }
3788 } else if (obj->is_Phi()) {
3789 Node* region = obj->in(0);
3790 // TODO make this more robust (see JDK-8231346)
3791 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3792 IfNode* iff = region->in(2)->in(0)->isa_If();
3793 if (iff != nullptr) {
3794 iff->is_flat_array_check(&_gvn, &array);
3795 }
3796 }
3797 }
3798 if (array != nullptr) {
3799 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3800 if (ary_t != nullptr) {
3801 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3802 // Casting array element to a non-inline-type, mark array as not null-free.
3803 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3804 replace_in_map(array, cast);
3805 array = cast;
3806 }
3807 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3808 // Casting array element to a non-flat-in-array type, mark array as not flat.
3809 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3810 replace_in_map(array, cast);
3811 array = cast;
3812 }
3813 }
3814 }
3815 }
3816
3817 if (!stopped() && !res->is_InlineType()) {
3818 res = record_profiled_receiver_for_speculation(res);
3819 if (toop->is_inlinetypeptr() && !maybe_larval) {
3820 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3821 res = vt;
3822 if (safe_for_replace) {
3823 replace_in_map(obj, vt);
3824 replace_in_map(not_null_obj, vt);
3825 replace_in_map(res, vt);
3826 }
3827 }
3828 }
3829 return res;
3830 }
3831
3832 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3833 // Load markword
3834 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3835 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3836 if (check_lock && !UseCompactObjectHeaders) {
3837 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3838 // Check if obj is locked
3839 Node* locked_bit = MakeConX(markWord::unlocked_value);
3840 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3841 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3842 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3843 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3844 _gvn.transform(iff);
3845 Node* locked_region = new RegionNode(3);
3846 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3847
3848 // Unlocked: Use bits from mark word
3849 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3850 mark_phi->init_req(1, mark);
3851
3852 // Locked: Load prototype header from klass
3853 set_control(_gvn.transform(new IfFalseNode(iff)));
3854 // Make loads control dependent to make sure they are only executed if array is locked
3855 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3856 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3857 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
3858 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3859
3860 locked_region->init_req(2, control());
3861 mark_phi->init_req(2, proto);
3862 set_control(_gvn.transform(locked_region));
3863 record_for_igvn(locked_region);
3864
3865 mark = mark_phi;
3866 }
3867
3868 // Now check if mark word bits are set
3869 Node* mask = MakeConX(mask_val);
3870 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3871 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3872 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3873 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3874 }
3875
3876 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3877 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3878 }
3879
3880 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3881 // We can't use immutable memory here because the mark word is mutable.
3882 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3883 // check is moved out of loops (mainly to enable loop unswitching).
3884 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3885 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3886 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3887 }
3888
3889 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3890 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3891 }
3892
3893 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3894 assert(vk->has_null_free_atomic_layout() || vk->has_null_free_non_atomic_layout(), "Can't be null-free and flat");
3895
3896 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3897 if (!vk->has_null_free_non_atomic_layout()) {
3898 return intcon(1); // Always atomic
3899 } else if (!vk->has_null_free_atomic_layout()) {
3900 return intcon(0); // Never atomic
3901 }
3902
3903 Node* array_klass = load_object_klass(array);
3904 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3905 Node* layout_kind_addr = basic_plus_adr(array_klass, array_klass, layout_kind_offset);
3906 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3907 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
3908 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3909 }
3910
3911 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3912 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3913 RegionNode* region = new RegionNode(3);
3914 Node* null_ctl = top();
3915 null_check_oop(val, &null_ctl);
3916 if (null_ctl != top()) {
3917 PreserveJVMState pjvms(this);
3918 set_control(null_ctl);
3919 {
3920 // Deoptimize if null-free array
3921 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3922 inc_sp(nargs);
3923 uncommon_trap(Deoptimization::Reason_null_check,
3924 Deoptimization::Action_none);
3925 }
3926 region->init_req(1, control());
3927 }
3928 region->init_req(2, control());
3929 set_control(_gvn.transform(region));
3930 record_for_igvn(region);
3931 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3932 // Since we were just successfully storing null, the array can't be null free.
3933 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3934 ary_t = ary_t->cast_to_not_null_free();
3935 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3936 if (safe_for_replace) {
3937 replace_in_map(ary, cast);
3938 }
3939 ary = cast;
3940 }
3941 return ary;
3942 }
3943
3944 //------------------------------next_monitor-----------------------------------
3945 // What number should be given to the next monitor?
3946 int GraphKit::next_monitor() {
3947 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3948 int next = current + C->sync_stack_slots();
3949 // Keep the toplevel high water mark current:
3950 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3951 return current;
3952 }
3953
3954 //------------------------------insert_mem_bar---------------------------------
3955 // Memory barrier to avoid floating things around
3956 // The membar serves as a pinch point between both control and all memory slices.
3957 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3958 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3959 mb->init_req(TypeFunc::Control, control());
3960 mb->init_req(TypeFunc::Memory, reset_memory());
3961 Node* membar = _gvn.transform(mb);
4055 lock->create_lock_counter(map()->jvms());
4056 increment_counter(lock->counter()->addr());
4057 }
4058 #endif
4059
4060 return flock;
4061 }
4062
4063
4064 //------------------------------shared_unlock----------------------------------
4065 // Emit unlocking code.
4066 void GraphKit::shared_unlock(Node* box, Node* obj) {
4067 // bci is either a monitorenter bc or InvocationEntryBci
4068 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4069 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4070
4071 if (stopped()) { // Dead monitor?
4072 map()->pop_monitor(); // Kill monitor from debug info
4073 return;
4074 }
4075 assert(!obj->is_InlineType(), "should not unlock on inline type");
4076
4077 // Memory barrier to avoid floating things down past the locked region
4078 insert_mem_bar(Op_MemBarReleaseLock);
4079
4080 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4081 UnlockNode *unlock = new UnlockNode(C, tf);
4082 #ifdef ASSERT
4083 unlock->set_dbg_jvms(sync_jvms());
4084 #endif
4085 uint raw_idx = Compile::AliasIdxRaw;
4086 unlock->init_req( TypeFunc::Control, control() );
4087 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4088 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4089 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4090 unlock->init_req( TypeFunc::ReturnAdr, top() );
4091
4092 unlock->init_req(TypeFunc::Parms + 0, obj);
4093 unlock->init_req(TypeFunc::Parms + 1, box);
4094 unlock = _gvn.transform(unlock)->as_Unlock();
4095
4096 Node* mem = reset_memory();
4097
4098 // unlock has no side-effects, sets few values
4099 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4100
4101 // Kill monitor from debug info
4102 map()->pop_monitor( );
4103 }
4104
4105 //-------------------------------get_layout_helper-----------------------------
4106 // If the given klass is a constant or known to be an array,
4107 // fetch the constant layout helper value into constant_value
4108 // and return null. Otherwise, load the non-constant
4109 // layout helper value, and return the node which represents it.
4110 // This two-faced routine is useful because allocation sites
4111 // almost always feature constant types.
4112 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4113 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4114 if (!StressReflectiveCode && klass_t != nullptr) {
4115 bool xklass = klass_t->klass_is_exact();
4116 bool can_be_flat = false;
4117 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4118 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4119 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4120 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4121 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4122 }
4123 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4124 jint lhelper;
4125 if (klass_t->is_flat()) {
4126 lhelper = ary_type->flat_layout_helper();
4127 } else if (klass_t->isa_aryklassptr()) {
4128 BasicType elem = ary_type->elem()->array_element_basic_type();
4129 if (is_reference_type(elem, true)) {
4130 elem = T_OBJECT;
4131 }
4132 lhelper = Klass::array_layout_helper(elem);
4133 } else {
4134 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4135 }
4136 if (lhelper != Klass::_lh_neutral_value) {
4137 constant_value = lhelper;
4138 return (Node*) nullptr;
4139 }
4140 }
4141 }
4142 constant_value = Klass::_lh_neutral_value; // put in a known value
4143 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
4144 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4145 }
4146
4147 // We just put in an allocate/initialize with a big raw-memory effect.
4148 // Hook selected additional alias categories on the initialization.
4149 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4150 MergeMemNode* init_in_merge,
4151 Node* init_out_raw) {
4152 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4153 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4154
4155 Node* prevmem = kit.memory(alias_idx);
4156 init_in_merge->set_memory_at(alias_idx, prevmem);
4157 if (init_out_raw != nullptr) {
4158 kit.set_memory(init_out_raw, alias_idx);
4159 }
4160 }
4161
4162 //---------------------------set_output_for_allocation-------------------------
4163 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4164 const TypeOopPtr* oop_type,
4165 bool deoptimize_on_exception) {
4166 int rawidx = Compile::AliasIdxRaw;
4167 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4168 add_safepoint_edges(alloc);
4169 Node* allocx = _gvn.transform(alloc);
4170 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4171 // create memory projection for i_o
4172 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4173 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4174
4175 // create a memory projection as for the normal control path
4176 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4177 set_memory(malloc, rawidx);
4178
4179 // a normal slow-call doesn't change i_o, but an allocation does
4180 // we create a separate i_o projection for the normal control path
4181 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4182 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4183
4184 // put in an initialization barrier
4185 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4186 rawoop)->as_Initialize();
4187 assert(alloc->initialization() == init, "2-way macro link must work");
4188 assert(init ->allocation() == alloc, "2-way macro link must work");
4189 {
4190 // Extract memory strands which may participate in the new object's
4191 // initialization, and source them from the new InitializeNode.
4192 // This will allow us to observe initializations when they occur,
4193 // and link them properly (as a group) to the InitializeNode.
4194 assert(init->in(InitializeNode::Memory) == malloc, "");
4195 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4196 init->set_req(InitializeNode::Memory, minit_in);
4197 record_for_igvn(minit_in); // fold it up later, if possible
4198 _gvn.set_type(minit_in, Type::MEMORY);
4199 Node* minit_out = memory(rawidx);
4200 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4201 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4202 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4203 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4204 // multiple projections as a result.
4205 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4206 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4207 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4208 if (oop_type->isa_aryptr()) {
4209 // Initially all flat array accesses share a single slice
4210 // but that changes after parsing. Prepare the memory graph so
4211 // it can optimize flat array accesses properly once they
4212 // don't share a single slice.
4213 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4214 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4215 int elemidx = C->get_alias_index(telemref);
4216 const TypePtr* alias_adr_type = C->get_adr_type(elemidx);
4217 if (alias_adr_type->is_flat()) {
4218 C->set_flat_accesses();
4219 }
4220 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, alias_adr_type)));
4221 } else if (oop_type->isa_instptr()) {
4222 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4223 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4224 ciField* field = ik->nonstatic_field_at(i);
4225 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4226 continue; // do not bother to track really large numbers of fields
4227 // Find (or create) the alias category for this field:
4228 int fieldidx = C->alias_type(field)->index();
4229 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4230 }
4231 }
4232 }
4233
4234 // Cast raw oop to the real thing...
4235 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4236 javaoop = _gvn.transform(javaoop);
4237 C->set_recent_alloc(control(), javaoop);
4238 assert(just_allocated_object(control()) == javaoop, "just allocated");
4239
4240 #ifdef ASSERT
4252 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4253 }
4254 }
4255 #endif //ASSERT
4256
4257 return javaoop;
4258 }
4259
4260 //---------------------------new_instance--------------------------------------
4261 // This routine takes a klass_node which may be constant (for a static type)
4262 // or may be non-constant (for reflective code). It will work equally well
4263 // for either, and the graph will fold nicely if the optimizer later reduces
4264 // the type to a constant.
4265 // The optional arguments are for specialized use by intrinsics:
4266 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4267 // - If 'return_size_val', report the total object size to the caller.
4268 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4269 Node* GraphKit::new_instance(Node* klass_node,
4270 Node* extra_slow_test,
4271 Node* *return_size_val,
4272 bool deoptimize_on_exception,
4273 InlineTypeNode* inline_type_node) {
4274 // Compute size in doublewords
4275 // The size is always an integral number of doublewords, represented
4276 // as a positive bytewise size stored in the klass's layout_helper.
4277 // The layout_helper also encodes (in a low bit) the need for a slow path.
4278 jint layout_con = Klass::_lh_neutral_value;
4279 Node* layout_val = get_layout_helper(klass_node, layout_con);
4280 bool layout_is_con = (layout_val == nullptr);
4281
4282 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4283 // Generate the initial go-slow test. It's either ALWAYS (return a
4284 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4285 // case) a computed value derived from the layout_helper.
4286 Node* initial_slow_test = nullptr;
4287 if (layout_is_con) {
4288 assert(!StressReflectiveCode, "stress mode does not use these paths");
4289 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4290 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4291 } else { // reflective case
4292 // This reflective path is used by Unsafe.allocateInstance.
4293 // (It may be stress-tested by specifying StressReflectiveCode.)
4294 // Basically, we want to get into the VM is there's an illegal argument.
4295 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4296 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4297 if (extra_slow_test != intcon(0)) {
4298 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4299 }
4300 // (Macro-expander will further convert this to a Bool, if necessary.)
4311
4312 // Clear the low bits to extract layout_helper_size_in_bytes:
4313 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4314 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4315 size = _gvn.transform( new AndXNode(size, mask) );
4316 }
4317 if (return_size_val != nullptr) {
4318 (*return_size_val) = size;
4319 }
4320
4321 // This is a precise notnull oop of the klass.
4322 // (Actually, it need not be precise if this is a reflective allocation.)
4323 // It's what we cast the result to.
4324 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4325 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4326 const TypeOopPtr* oop_type = tklass->as_instance_type();
4327
4328 // Now generate allocation code
4329
4330 // The entire memory state is needed for slow path of the allocation
4331 // since GC and deoptimization can happen.
4332 Node *mem = reset_memory();
4333 set_all_memory(mem); // Create new memory state
4334
4335 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4336 control(), mem, i_o(),
4337 size, klass_node,
4338 initial_slow_test, inline_type_node);
4339
4340 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4341 }
4342
4343 //-------------------------------new_array-------------------------------------
4344 // helper for newarray and anewarray
4345 // The 'length' parameter is (obviously) the length of the array.
4346 // The optional arguments are for specialized use by intrinsics:
4347 // - If 'return_size_val', report the non-padded array size (sum of header size
4348 // and array body) to the caller.
4349 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4350 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4351 Node* length, // number of array elements
4352 int nargs, // number of arguments to push back for uncommon trap
4353 Node* *return_size_val,
4354 bool deoptimize_on_exception,
4355 Node* init_val) {
4356 jint layout_con = Klass::_lh_neutral_value;
4357 Node* layout_val = get_layout_helper(klass_node, layout_con);
4358 bool layout_is_con = (layout_val == nullptr);
4359
4360 if (!layout_is_con && !StressReflectiveCode &&
4361 !too_many_traps(Deoptimization::Reason_class_check)) {
4362 // This is a reflective array creation site.
4363 // Optimistically assume that it is a subtype of Object[],
4364 // so that we can fold up all the address arithmetic.
4365 layout_con = Klass::array_layout_helper(T_OBJECT);
4366 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4367 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4368 { BuildCutout unless(this, bol_lh, PROB_MAX);
4369 inc_sp(nargs);
4370 uncommon_trap(Deoptimization::Reason_class_check,
4371 Deoptimization::Action_maybe_recompile);
4372 }
4373 layout_val = nullptr;
4374 layout_is_con = true;
4375 }
4376
4377 // Generate the initial go-slow test. Make sure we do not overflow
4378 // if length is huge (near 2Gig) or negative! We do not need
4379 // exact double-words here, just a close approximation of needed
4380 // double-words. We can't add any offset or rounding bits, lest we
4381 // take a size -1 of bytes and make it positive. Use an unsigned
4382 // compare, so negative sizes look hugely positive.
4383 int fast_size_limit = FastAllocateSizeLimit;
4384 if (layout_is_con) {
4385 assert(!StressReflectiveCode, "stress mode does not use these paths");
4386 // Increase the size limit if we have exact knowledge of array type.
4387 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4388 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4389 }
4390
4391 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4392 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4393
4394 // --- Size Computation ---
4395 // array_size = round_to_heap(array_header + (length << elem_shift));
4396 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4397 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4398 // The rounding mask is strength-reduced, if possible.
4399 int round_mask = MinObjAlignmentInBytes - 1;
4400 Node* header_size = nullptr;
4401 // (T_BYTE has the weakest alignment and size restrictions...)
4402 if (layout_is_con) {
4403 int hsize = Klass::layout_helper_header_size(layout_con);
4404 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4405 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4406 if ((round_mask & ~right_n_bits(eshift)) == 0)
4407 round_mask = 0; // strength-reduce it if it goes away completely
4408 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4409 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4410 assert(header_size_min <= hsize, "generic minimum is smallest");
4411 header_size = intcon(hsize);
4412 } else {
4413 Node* hss = intcon(Klass::_lh_header_size_shift);
4414 Node* hsm = intcon(Klass::_lh_header_size_mask);
4415 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4416 header_size = _gvn.transform(new AndINode(header_size, hsm));
4417 }
4418
4419 Node* elem_shift = nullptr;
4420 if (layout_is_con) {
4421 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4422 if (eshift != 0)
4423 elem_shift = intcon(eshift);
4424 } else {
4425 // There is no need to mask or shift this value.
4426 // The semantics of LShiftINode include an implicit mask to 0x1F.
4427 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4428 elem_shift = layout_val;
4477 }
4478 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4479
4480 if (return_size_val != nullptr) {
4481 // This is the size
4482 (*return_size_val) = non_rounded_size;
4483 }
4484
4485 Node* size = non_rounded_size;
4486 if (round_mask != 0) {
4487 Node* mask1 = MakeConX(round_mask);
4488 size = _gvn.transform(new AddXNode(size, mask1));
4489 Node* mask2 = MakeConX(~round_mask);
4490 size = _gvn.transform(new AndXNode(size, mask2));
4491 }
4492 // else if round_mask == 0, the size computation is self-rounding
4493
4494 // Now generate allocation code
4495
4496 // The entire memory state is needed for slow path of the allocation
4497 // since GC and deoptimization can happen.
4498 Node *mem = reset_memory();
4499 set_all_memory(mem); // Create new memory state
4500
4501 if (initial_slow_test->is_Bool()) {
4502 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4503 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4504 }
4505
4506 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4507 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4508
4509 Node* raw_init_value = nullptr;
4510 if (init_val != nullptr) {
4511 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4512 if (ary_type->is_flat()) {
4513 initial_slow_test = intcon(1);
4514 }
4515
4516 if (UseCompressedOops) {
4517 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4518 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4519 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4520 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4521 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4522 } else {
4523 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4524 }
4525 }
4526
4527 Node* valid_length_test = _gvn.intcon(1);
4528 if (ary_type->isa_aryptr()) {
4529 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4530 jint max = TypeAryPtr::max_array_length(bt);
4531 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4532 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4533 }
4534
4535 // Create the AllocateArrayNode and its result projections
4536 AllocateArrayNode* alloc
4537 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4538 control(), mem, i_o(),
4539 size, klass_node,
4540 initial_slow_test,
4541 length, valid_length_test,
4542 init_val, raw_init_value);
4543 // Cast to correct type. Note that the klass_node may be constant or not,
4544 // and in the latter case the actual array type will be inexact also.
4545 // (This happens via a non-constant argument to inline_native_newArray.)
4546 // In any case, the value of klass_node provides the desired array type.
4547 const TypeInt* length_type = _gvn.find_int_type(length);
4548 if (ary_type->isa_aryptr() && length_type != nullptr) {
4549 // Try to get a better type than POS for the size
4550 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4551 }
4552
4553 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4554
4555 array_ideal_length(alloc, ary_type, true);
4556 return javaoop;
4557 }
4558
4559 // The following "Ideal_foo" functions are placed here because they recognize
4560 // the graph shapes created by the functions immediately above.
4561
4562 //---------------------------Ideal_allocation----------------------------------
4657 void GraphKit::add_parse_predicates(int nargs) {
4658 if (ShortRunningLongLoop) {
4659 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4660 // walking up from the loop.
4661 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4662 }
4663 if (UseLoopPredicate) {
4664 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4665 if (UseProfiledLoopPredicate) {
4666 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4667 }
4668 }
4669 if (UseAutoVectorizationPredicate) {
4670 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4671 }
4672 // Loop Limit Check Predicate should be near the loop.
4673 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4674 }
4675
4676 void GraphKit::sync_kit(IdealKit& ideal) {
4677 reset_memory();
4678 set_all_memory(ideal.merged_memory());
4679 set_i_o(ideal.i_o());
4680 set_control(ideal.ctrl());
4681 }
4682
4683 void GraphKit::final_sync(IdealKit& ideal) {
4684 // Final sync IdealKit and graphKit.
4685 sync_kit(ideal);
4686 }
4687
4688 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4689 Node* len = load_array_length(load_String_value(str, set_ctrl));
4690 Node* coder = load_String_coder(str, set_ctrl);
4691 // Divide length by 2 if coder is UTF16
4692 return _gvn.transform(new RShiftINode(len, coder));
4693 }
4694
4695 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4696 int value_offset = java_lang_String::value_offset();
4697 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4698 false, nullptr, Type::Offset(0));
4699 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4700 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4701 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4702 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4703 Node* p = basic_plus_adr(str, str, value_offset);
4704 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4705 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4706 return load;
4707 }
4708
4709 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4710 if (!CompactStrings) {
4711 return intcon(java_lang_String::CODER_UTF16);
4712 }
4713 int coder_offset = java_lang_String::coder_offset();
4714 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4715 false, nullptr, Type::Offset(0));
4716 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4717
4718 Node* p = basic_plus_adr(str, str, coder_offset);
4719 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4720 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4721 return load;
4722 }
4723
4724 void GraphKit::store_String_value(Node* str, Node* value) {
4725 int value_offset = java_lang_String::value_offset();
4726 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4727 false, nullptr, Type::Offset(0));
4728 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4729
4730 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4731 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4732 }
4733
4734 void GraphKit::store_String_coder(Node* str, Node* value) {
4735 int coder_offset = java_lang_String::coder_offset();
4736 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4737 false, nullptr, Type::Offset(0));
4738 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4739
4740 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4741 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4742 }
4743
4744 // Capture src and dst memory state with a MergeMemNode
4745 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4746 if (src_type == dst_type) {
4747 // Types are equal, we don't need a MergeMemNode
4748 return memory(src_type);
4749 }
4750 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4751 record_for_igvn(merge); // fold it up later, if possible
4752 int src_idx = C->get_alias_index(src_type);
4753 int dst_idx = C->get_alias_index(dst_type);
4754 merge->set_memory_at(src_idx, memory(src_idx));
4755 merge->set_memory_at(dst_idx, memory(dst_idx));
4756 return merge;
4757 }
4830 i_char->init_req(2, AddI(i_char, intcon(2)));
4831
4832 set_control(IfFalse(iff));
4833 set_memory(st, TypeAryPtr::BYTES);
4834 }
4835
4836 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4837 if (!field->is_constant()) {
4838 return nullptr; // Field not marked as constant.
4839 }
4840 ciInstance* holder = nullptr;
4841 if (!field->is_static()) {
4842 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4843 if (const_oop != nullptr && const_oop->is_instance()) {
4844 holder = const_oop->as_instance();
4845 }
4846 }
4847 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4848 /*is_unsigned_load=*/false);
4849 if (con_type != nullptr) {
4850 Node* con = makecon(con_type);
4851 if (field->type()->is_inlinetype()) {
4852 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4853 } else if (con_type->is_inlinetypeptr()) {
4854 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4855 }
4856 return con;
4857 }
4858 return nullptr;
4859 }
4860
4861 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4862 const Type* obj_type = obj->bottom_type();
4863 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4864 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4865 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4866 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4867 obj = casted_obj;
4868 }
4869 if (sig_type->is_inlinetypeptr()) {
4870 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4871 }
4872 return obj;
4873 }
|